Karl的博客

【论文列表】2025年Object Pose Estimation主要会议论文列表

2026-07-02T14:29:51.000Z

CCF A

CVPR

Any6D: Model-free 6D Pose Estimation of Novel Objects
- 分类：任意物体方法
- 论文：https://openaccess.thecvf.com/content/CVPR2025/html/Lee_Any6D_Model-free_6D_Pose_Estimation_of_Novel_Objects_CVPR_2025_paper.html
- 代码：https://github.com/taeyeopl/Any6D
- 摘要：We introduce Any6D, a model-free framework for 6D object poseestimation that requires only a single RGB-D anchor image to estimateboth the 6D pose and size of unknown objects in novel scenes. Unlikeexisting methods that rely on textured 3D models or multiple viewpoints,Any6D leverages a joint object alignment process to enhance 2D-3Dalignment and metric scale estimation for improved pose accuracy. Ourapproach integrates a render-and-compare strategy to generate and refinepose hypotheses, enabling robust performance in scenarios withocclusions, non-overlapping views, diverse lighting conditions, andlarge cross-environment variations. We evaluate our method on fivechallenging datasets: REAL275, Toyota-Light, HO3D, YCBINEOAT, and LM-O,demonstrating its effectiveness in significantly outperformingstate-of-the-art methods for novel object pose estimation. Project page:https://taeyeop.com/any6d.
CAP-Net: A Unified Network for 6D Pose and Size Estimation ofCategorical Articulated Parts from a Single RGB-D Image
- 分类：类别级铰接部件方法
- 论文：https://openaccess.thecvf.com/content/CVPR2025/html/Huang_CAP-Net_A_Unified_Network_for_6D_Pose_and_Size_Estimation_CVPR_2025_paper.html
- 代码：https://github.com/ShaneHuangHZ/CAPNet
- 摘要：This paper tackles category-level pose estimation ofarticulated objects in robotic manipulation tasks and introduces a newbenchmark dataset. While recent methods estimate part poses and sizes atthe category level, they often rely on geometric cues and complexmulti-stage pipelines that first segment parts from the point cloud,followed by Normalized Part Coordinate Space (NPCS) estimation for 6Dposes. These approaches overlook dense semantic cues from RGB images,leading to suboptimal accuracy, particularly for objects with smallparts. To address these limitations, we propose a single-stage Network,CAP-Net, for estimating the 6D poses and sizes of CategoricalArticulated Parts. This method combines RGB-D features to generateinstance segmentation and NPCS representations for each part in anend-to-end manner. CAP-Net uses a unified network to simultaneouslypredict point-wise class labels, centroid offsets, and NPCS maps. Aclustering algorithm then groups points of the same predicted classbased on their estimated centroid distances to isolate each part.Finally, the NPCS region of each part is aligned with the point cloud torecover its final pose and size. To bridge the sim-to-real domain gap,we introduce the RGBD-Art dataset, the largest RGB-D articulated datasetto date, featuring photorealistic RGB images and depth noise simulatedfrom real sensors. Experimental evaluations on the RGBD-Art datasetdemonstrate that our method significantly outperforms thestate-of-the-art approach. Real-world deployments of our model inrobotic tasks underscore its robustness and exceptional sim-to-realtransfer capabilities, confirming its substantial practical utility. Ourdataset, code and pre-trained models are available on the projectpage.
Co-op: Correspondence-based Novel Object Pose Estimation
- 分类：任意物体方法
- 论文：https://openaccess.thecvf.com/content/CVPR2025/html/Moon_Co-op_Correspondence-based_Novel_Object_Pose_Estimation_CVPR_2025_paper.html
- 代码：代码未开源
- 摘要：We propose Co-op, a novel method for accurately and robustlyestimating the 6DoF pose of objects unseen during training from a singleRGB image. Our method requires only the CAD model of the target objectand can precisely estimate its pose without any additional fine-tuning.While existing model-based methods suffer from inefficiency due to usinga large number of templates, our method enables fast and accurateestimation with a small number of templates. This improvement isachieved by finding semi-dense correspondences between the input imageand the pre-rendered templates. Our method achieves stronggeneralization performance by leveraging a hybrid representation thatcombines patch-level classification and offset regression. Additionally,our pose refinement model estimates probabilistic flow between the inputimage and the rendered image, refining the initial estimate to anaccurate pose using a differentiable PnP layer. We demonstrate that ourmethod not only estimates object poses rapidly but also outperformsexisting methods by a large margin on the seven core datasets of the BOPChallenge, achieving state-of-the-art accuracy.
CRISP: Object Pose and Shape Estimation with Test-Time Adaptation
- 分类：任意物体方法
- 论文：https://openaccess.thecvf.com/content/CVPR2025/html/Shi_CRISP_Object_Pose_and_Shape_Estimation_with_Test-Time_Adaptation_CVPR_2025_paper.html
- 代码：https://github.com/MIT-SPARK/CRISP
- 摘要：We consider the problem of estimating object pose and shapefrom an RGB-D image. Our first contribution is to introduce CRISP, acategory-agnostic object pose and shape estimation pipeline. Thepipeline implements an encoder-decoder model for shape estimation. Ituses FiLM-conditioning for implicit shape reconstruction and a DPT-basednetwork for estimating pose-normalized points for pose estimation. As asecond contribution, we propose an optimization-based pose and shapecorrector that can correct estimation errors caused by a domain gap.Observing that the shape decoder is well behaved in the convex hull ofknown shapes, we approximate the shape decoder with an active shapemodel, and show that this reduces the shape correction problem to aconstrained linear least squares problem, which can be solvedefficiently by an interior point algorithm. Third, we introduce aself-training pipeline to perform self-supervised domain adaptation ofCRISP. The self-training is based on a correct-and-certify approach,which leverages the corrector to generate pseudo-labels at test time,and uses them to self-train CRISP. We demonstrate CRISP (and theself-training) on YCBV, SPE3R, and NOCS datasets. CRISP shows highperformance on all the datasets. Moreover, our self-training is capableof bridging a large domain gap. Finally, CRISP also shows an ability togeneralize to unseen objects. Code, pre-trained models and videos ofsample results are available on the project webpage.
GCE-Pose: Global Context Enhancement for Category-level Object PoseEstimation
- 分类：类别级方法
- 论文：https://openaccess.thecvf.com/content/CVPR2025/html/Li_GCE-Pose_Global_Context_Enhancement_for_Category-level_Object_Pose_Estimation_CVPR_2025_paper.html
- 代码：代码未开源
- 摘要：A key challenge in model-free category-level pose estimationis the extraction of contextual object features that generalize acrossvarying instances within a specific category. Recent approaches leveragefoundational features to capture semantic and geometry cues from data.However, these approaches fail under partial visibility. We overcomethis with a first-complete-then-aggregate strategy for featureextraction utilizing class priors. In this paper, we present GCE-Pose, amethod that enhances pose estimation for novel instances by integratingcategory-level global context prior. GCE-Pose performs semantic shapereconstruction with a proposed Semantic Shape Reconstruction (SSR)module. Given an unseen partial RGB-D object instance, our SSR modulereconstructs the instance’s global geometry and semantics by deformingcategory-specific 3D semantic prototypes through a learned deep LinearShape Model. We further introduce a Global Context Enhanced (GCE)feature fusion module that effectively fuses features from partial RGB-Dobservations and the reconstructed global context. Extensive experimentsvalidate the impact of our global context prior and the effectiveness ofthe GCE fusion module, demonstrating that GCE-Pose significantlyoutperforms existing methods on challenging real-world datasetsHouse-Cat6D and NOCS-REAL275. Our project page is available at https://colin-de.github.io/GCE-Pose/.
GIVEPose: Gradual Intra-class Variation Elimination for RGB-basedCategory-Level Object Pose Estimation
- 分类：类别级方法
- 论文：https://openaccess.thecvf.com/content/CVPR2025/html/Huang_GIVEPose_Gradual_Intra-class_Variation_Elimination_for_RGB-based_Category-Level_Object_Pose_CVPR_2025_paper.html
- 代码：https://github.com/ziqin-h/GIVEPose
- 摘要：Recent advances in RGBD-based category-level object poseestimation have been limited by their reliance on precise depthinformation, restricting their broader applicability. In response,RGB-based methods have been developed. Among these methods,geometry-guided pose regression that originated from instance-leveltasks has demonstrated strong performance. However, we argue that theNOCS map is an inadequate intermediate representation forgeometry-guided pose regression method, as its many-to-onecorrespondence with category-level pose introduces redundantinstance-specific information, resulting in suboptimal results. Thispaper identifies the intra-class variation problem inherent in poseregression based solely on the NOCS map and proposes the Intra-classVariation-Free Consensus (IVFC) map, a novel coordinate representationgenerated from the category-level consensus model. By leveraging thecomplementary strengths of the NOCS map and the IVFC map, we introduceGIVEPose, a framework that implements Gradual Intra-class VariationElimination for category-level object pose estimation. Extensiveevaluations on both synthetic and real-world datasets demonstrate thatGIVEPose significantly outperforms existing state-of-the-art RGB-basedapproaches, achieving substantial improvements in category-level objectpose estimation. Our code is available at https://github.com/ziqin-h/GIVEPose.
iG-6DoF: Model-free 6DoF Pose Estimation for Unseen Object viaIterative 3D Gaussian Splatting
- 分类：任意物体方法
- 论文：https://openaccess.thecvf.com/content/CVPR2025/html/Cao_iG-6DoF_Model-free_6DoF_Pose_Estimation_for_Unseen_Object_via_Iterative_CVPR_2025_paper.html
- 代码：代码未开源
- 摘要：Traditional methods in pose estimation often rely on precise3D models or additional data such as depth and normals, limiting theirgeneralization, especially when objects undergo large translations orrotations. We propose iG-6DoF, a novel model-free 6D pose estimationmethod using iterative 3D Gaussian Splatting to estimate the pose ofunseen objects. We first estimates an initial pose by leveragingmulti-scale data augmentation and the rotation-equivariant features tocreate a better pose hypothesis from a set of candidates. Then, wepropose an iterative 3DGS approach through iteratively rendering andcomparing the rendered image with the input image to furtherprogressively improve pose estimation accuracy. The proposed methodconsists of an object detector, a multi-scale rotation-equivariantfeature based initial pose estimator, and a coarse-to-fine pose refiner.Such combination allows our method to focus on the target object in acomplex scene dealing with large movement and weak textures. Our methodachieves state-of-the-art results on the LINEMOD, OnePose-LowTexture,GenMOP datasets and our self-captured data, demonstrating its stronggeneralization to unseen objects and robustness across variousscenes.
Leveraging Global Stereo Consistency for Category-Level Shape and 6DPose Estimation from Stereo Images
- 分类：类别级方法
- 论文：https://openaccess.thecvf.com/content/CVPR2025/html/Qiu_Leveraging_Global_Stereo_Consistency_for_Category-Level_Shape_and_6D_Pose_CVPR_2025_paper.html
- 代码：代码未开源
- 摘要：Stereo-based category-level shape and 6D pose estimationmethods have the potential to generalize to a wider range of materialsthan RGBD methods, which often suffer from depth measurement errors.However, without explicit depth from two views, parameters to beestimated can become inherently entangled, negatively impactingperformance. To address this, we propose a method that leverages globalstereo consistency to constrain optimization directions and mitigateparameter entanglement. We first estimate an intra-category occupancyfield to represent a unified shape across views, ensuring consistencyand preventing shape ambiguity. Through a divide-and-conquer approachwithin global shape fitting, we fit this shape to stereo images toobtain the pose, iteratively rendering normalized depth maps andexchanging information across views. This approach improves convergencetoward the correct pose and scale. We validated our method on bothdepth-friendly and depth-challenging materials using our S-RGBD datasetand the TOD benchmark. Our method surpasses RGBD methods on challengingobjects and performs comparably on depth-friendly ones. Ablation studiesconfirm the effectiveness of each component.
ONDA-Pose: Occlusion-Aware Neural Domain Adaptation forSelf-Supervised 6D Object Pose Estimation
- 分类：实例级方法
- 论文：https://openaccess.thecvf.com/content/CVPR2025/html/Tan_ONDA-Pose_Occlusion-Aware_Neural_Domain_Adaptation_for_Self-Supervised_6D_Object_Pose_CVPR_2025_paper.html
- 代码：https://github.com/tan-tao11/ONDA-Pose
- 摘要：Self-supervised 6D object pose estimation has receivedincreasing attention in computer vision recently. Some typical works inliterature attempt to translate the synthetic images with object poselabels generated by object CAD models into the real domain, and then usethe translated data for training. However, their performance isgenerally limited, since (i) there still exists a domain gap between thetranslated images and the real images and (ii) the translated images cannot sufficiently reflect occlusions that exist in many real images. Toaddress these problems, we propose an Occlusion-Aware Neural DomainAdaptation method for self-supervised 6D object Pose estimation, calledONDA-Pose. The proposed method comprises three main steps. Firstly, byutilizing both the training real images without pose labels and a CADmodel, we explore a CAD-like radiance field for rendering correspondingsynthetic images that have similar textures to those generated by theCAD model. Then, a backbone pose estimator trained on the synthetic datais employed to provide initial pose estimations for the synthetic imagesrendered from the CAD-like radiance field, and the initial object posesare refined by a global object pose refiner to generate pseudo objectpose labels. Finally, the backbone pose estimator is furtherself-supervised as the final pose estimator by jointly utilizing thereal images with pseudo object pose labels and the synthetic imagesrendered from the CAD-like radiance field. Experimental results on threepublic datasets demonstrate that ONDA-Pose significantly outperforms thecomparative state-of-the-art methods in most cases.
One2Any: One-Reference 6D Pose Estimation for Any Object
- 分类：任意物体方法
- 论文：https://openaccess.thecvf.com/content/CVPR2025/html/Liu_One2Any_One-Reference_6D_Pose_Estimation_for_Any_Object_CVPR_2025_paper.html
- 代码：https://github.com/lmy1001/One2Any
- 摘要：6D object pose estimation remains challenging for manyapplications due to dependencies on complete 3D models, multi-viewimages, or training limited to specific object categories. Theserequirements make generalization to novel objects difficult for whichneither 3D models nor multi-view images may be available. To addressthis, we propose a novel method One2Any that estimates the relative6-degrees of freedom (DOF) object pose using only a singlereference-single query RGB-D image, without prior knowledge of its 3Dmodel, multi-view data, or category constraints. We treat object poseestimation as an encoding-decoding process: first, we obtain acomprehensive Reference Object Pose Embedding (ROPE) that encodes anobject’s shape, orientation, and texture from a single reference view.Using this embedding, a U-Net-based pose decoding module producesReference Object Coordinate (ROC) for new views, enabling fast andaccurate pose estimation. This simple encoding-decoding framework allowsour model to be trained on any pair-wise pose data, enabling large-scaletraining and demonstrating great scalability. Experiments on multiplebenchmark datasets demonstrate that our model generalizes well to novelobjects, achieving state-of-the-art accuracy and robustness evenrivaling methods that require multi-view or CAD inputs, at a fraction ofcompute. Code is available at https://github.com/lmy1001/One2Any.
Pos3R: 6D Pose Estimation for Unseen Objects Made Easy
- 分类：任意物体方法
- 论文：https://openaccess.thecvf.com/content/CVPR2025/html/Deng_Pos3R_6D_Pose_Estimation_for_Unseen_Objects_Made_Easy_CVPR_2025_paper.html
- 代码：代码未开源
- 摘要：Foundation models have significantly reduced the need fortask-specific training, while also enhancing generalizability. However,state-of-the-art 6D pose estimators either require further training withpose supervision or neglect advances obtainable from 3D foundationmodels. The latter is a missed opportunity, since these models arebetter equipped to predict 3D-consistent features, which are ofsignificant utility for the pose estimation task. To address this gap,we propose Pos3R, a method for estimating the 6D pose of any object froma single RGB image, making extensive use of a 3D reconstructionfoundation model and requiring no additional training. We identifytemplate selection as a particular bottleneck for existing methods thatis significantly alleviated by the use of a 3D model, which can moreeasily distinguish between template poses than a 2D model. Despite itssimplicity, Pos3R achieves competitive performance on the Benchmark for6D Object Pose Estimation (BOP), matching or surpassing existingrefinement-free methods. Additionally, Pos3R integrates seamlessly withrender-and-compare refinement techniques, demonstrating adaptability forhigh-precision applications.
RefPose: Leveraging Reference Geometric Correspondences for Accurate6D Pose Estimation of Unseen Objects
- 分类：任意物体方法
- 论文：https://openaccess.thecvf.com/content/CVPR2025/html/Kim_RefPose_Leveraging_Reference_Geometric_Correspondences_for_Accurate_6D_Pose_Estimation_CVPR_2025_paper.html
- 代码：代码未开源
- 摘要：Estimating the 6D pose of unseen objects from monocular RGBimages remains a challenging problem, especially due to the lack ofprior object-specific knowledge. To tackle this issue, we proposeRefPose, an innovative approach to object pose estimation that leveragesa reference image and geometric correspondence as guidance. RefPosefirst predicts an initial pose by using object templates to render thereference image and establish the geometric correspondence needed forthe refinement stage. During the refinement stage, RefPose estimates thegeometric correspondence of the query image based on the generatedreferences and iteratively refines the pose through a render-and-compareapproach. To enhance this estimation, we introduce a correlationvolume-guided attention mechanism that effectively captures correlationsbetween the query and reference images. Unlike traditional methods thatdepend on pre-defined object models, RefPose dynamically adapts to newobject shapes by leveraging a reference image and geometriccorrespondence. This results in robust performance across previouslyunseen objects. Extensive evaluation on the BOP benchmark datasets showsthat RefPose achieves state-of-the-art results while maintaining acompetitive runtime.
Rethinking Correspondence-based Category-Level Object PoseEstimation
- 分类：类别级方法
- 论文：https://openaccess.thecvf.com/content/CVPR2025/html/Ren_Rethinking_Correspondence-based_Category-Level_Object_Pose_Estimation_CVPR_2025_paper.html
- 代码：代码未开源
- 摘要：Category-level object pose estimation aims to determine thepose and size of arbitrary objects within given categories. Existingtwo-stage correspondence-based methods first establish correspondencesbetween camera and object coordinates, and then acquire the object poseusing a pose fitting algorithm. In this paper, we conduct acomprehensive analysis of this paradigm and introduce two crucialessentials: 1) shape-sensitive and pose-invariant feature extraction foraccurate correspondence prediction, and 2) outlier correspondenceremoval for robust pose fitting. Based on these insights, we propose asimple yet effective correspondence-based method called SpotPose, whichincludes two stages. During the correspondence prediction stage,pose-invariant geometric structure of objects is thoroughly exploited tofacilitate shape-sensitive holistic interaction among keypoint-wisefeatures. During the pose fitting stage, outlier scores ofcorrespondences are explicitly predicted to facilitate efficientidentification and removal of outliers. Experimental results onCAMERA25, REAL275 and HouseCat6D benchmarks demonstrate that theproposed SpotPose outperforms state-of-the-art approaches by a largemargin.
SCFlow2: Plug-and-Play Object Pose Refiner with Shape-ConstraintScene Flow
- 分类：任意物体优化方法
- 论文：https://openaccess.thecvf.com/content/CVPR2025/html/Wang_SCFlow2_Plug-and-Play_Object_Pose_Refiner_with_Shape-Constraint_Scene_Flow_CVPR_2025_paper.html
- 代码：https://github.com/W-QY/SCFlow2
- 摘要：We introduce SCFlow2, a plug-and-play refinement framework for6D object pose estimation. Most recent 6D object pose methods rely onrefinement to get accurate results. However, most existing refinementmethods either suffer from noises in establishing correspondences, orrely on retraining for novel objects. SCFlow2 is based on the SCFlowmodel designed for refinement with shape constraint, but formulates theadditional depth as a regularization in the iteration via 3D scene flowfor RGBD frames. The key design of SCFlow2 is an introduction ofgeometry constraints into the training of recurrent matching network, bycombining the rigid-motion embeddings in 3D scene flow and 3D shapeprior of the target. We train SCFlow2 on a combination of datasetObjaverse, GSO and ShapeNet, and evaluate on BOP datasets with novelobjects. After using our method as a post-processing, moststate-of-the-art methods produce significantly better results, withoutany retraining or fine-tuning. The source code is available at https://scflow2.github.io.
Structure-Aware Correspondence Learning for Relative Pose Estimation
- 分类：任意物体方法
- 论文：https://openaccess.thecvf.com/content/CVPR2025/html/Chen_Structure-Aware_Correspondence_Learning_for_Relative_Pose_Estimation_CVPR_2025_paper.html
- 代码：https://github.com/Cyhhzo02/SAC-Pose-code
- 摘要：Relative pose estimation provides a promising way forachieving object-agnostic pose estimation. Despite the success ofexisting 3D correspondence-based methods, the reliance on explicitfeature matching suffers from small overlaps in visible regions andunreliable feature estimation for invisible regions. Inspired by humans’ability to assemble two object parts that have small or no overlappingregions by considering object structure, we propose a novelStructure-Aware Correspondence Learning method for Relative PoseEstimation, which consists of two key modules. First, a structure-awarekeypoint extraction module is designed to locate a set of keypoints thatcan represent the structure of objects with different shapes andappearance, under the guidance of a keypoint based image reconstructionloss. Second, a structure-aware correspondence estimation module isdesigned to model the intra-image and inter-image relationships betweenkeypoints to extract structure-aware features for correspondenceestimation. By jointly leveraging these two modules, the proposed methodcan naturally estimate 3D-3D correspondences for unseen objects withoutexplicit feature matching for precise relative pose estimation.Experimental results on the CO3D, Objaverse and LineMOD datasetsdemonstrate that the proposed method significantly outperforms priormethods, i.e., with 5.7° reduction in mean angular error on the CO3Ddataset.
UA-Pose: Uncertainty-Aware 6D Object Pose Estimation and OnlineObject Completion with Partial References
- 分类：任意物体方法
- 论文：https://openaccess.thecvf.com/content/CVPR2025/html/Li_UA-Pose_Uncertainty-Aware_6D_Object_Pose_Estimation_and_Online_Object_Completion_CVPR_2025_paper.html
- 代码：代码未开源
- 摘要：6D object pose estimation has shown strong generalizability tonovel objects. However, existing methods often require either acomplete, well-reconstructed 3D model or numerous reference images thatfully cover the object. Estimating 6D poses from partial references,which capture only fragments of an object’s appearance and geometry,remains challenging. To address this, we propose UA-Pose, anuncertainty-aware approach for 6D object pose estimation and onlineobject completion specifically designed for partial references. Weassume access to either (1) a limited set of RGBD images with knownposes or (2) a single 2D image. For the first case, we initialize apartial object 3D model based on the provided images and poses, whilefor the second, we use image-to-3D techniques to generate an initialobject 3D model. Our method integrates uncertainty into the incomplete3D model, distinguishing between seen and unseen regions. Thisuncertainty enables confidence assessment in pose estimation and guidesan uncertainty-aware sampling strategy for online object completion,enhancing robustness in pose estimation accuracy and improving objectcompleteness. We evaluate our method on the YCB-Video, YCBInEOAT, andHO3D datasets, including RGBD sequences of YCB objects manipulated byrobots and human hands. Experimental results demonstrate significantperformance improvements over existing methods, particularly when objectobservations are incomplete or partially captured. Project page: https://minfenli.github.io/UA-Pose/.
UNOPose: Unseen Object Pose Estimation with an Unposed RGB-DReference Image
- 分类：任意物体方法
- 论文：https://openaccess.thecvf.com/content/CVPR2025/html/Liu_UNOPose_Unseen_Object_Pose_Estimation_with_an_Unposed_RGB-D_Reference_CVPR_2025_paper.html
- 代码：https://github.com/shanice-l/UNOPose
- 摘要：Unseen object pose estimation methods often rely on CAD modelsor multiple reference views, making the onboarding stage costly. Tosimplify reference acquisition, we aim to estimate the unseen object’spose through a single unposed RGB-D reference image. While previousworks leverage reference images as pose anchors to limit the range ofrelative pose, our scenario presents significant challenges since therelative transformation could vary across the entire SE(3) space.Moreover, factors like occlusion, sensor noise, and extreme geometrycould result in low viewpoint overlap. To address these challenges, wepresent a novel approach and benchmark, termed UNOPose, for UNseenOne-reference-based object Pose estimation. Building upon acoarse-to-fine paradigm, UNOPose constructs an SE(3)-invariant referenceframe to standardize object representation despite pose and sizevariations. To alleviate small overlap across viewpoints, we recalibratethe weight of each correspondence based on its predicted likelihood ofbeing within the overlapping region. Evaluated on our proposed benchmarkbased on the BOP Challenge, UNOPose demonstrates superior performance,significantly outperforming traditional and learning-based methods inthe one-reference setting and remaining competitive with CAD-model-basedmethods. The code and dataset are available atgithub.com/shanice-l/UNOPose.

NeurIPS

RFMPose: Generative Category-level Object Pose Estimation viaRiemannian Flow Matching
- 分类：类别级方法
- 论文：https://neurips.cc/virtual/2025/poster/119779
- 代码：https://github.com/shabiouyang/RMFPose
- 摘要：We introduce RFMPose, a novel generative framework forcategory-level 6D object pose estimation that learns deterministic posetrajectories through Riemannian Flow Matching (RFM). Existingdiscriminative approaches struggle with multi-hypothesis predictions(e.g., symmetry ambiguities) and often require specialized networkarchitectures. RFMPose advances this paradigm through three keyinnovations: (1) Ensuring geometric consistency via geodesicinterpolation on Riemannian manifolds combined with bi-invariant metricconstraints; (2) Alleviating symmetry-induced ambiguities throughRiemannian Optimal Transport for probability mass redistribution withoutad-hoc design; (3) Enabling end-to-end likelihood estimation throughHutchinson trace approximation, thereby eliminating auxiliary modeldependencies. Extensive experiments on the Omni6DPose demonstratestate-of-the-art performance of the proposed method, with significantimprovements of +4.1 in IoU25 and +2.4 in 5°2cm metrics compared toprior generative approaches. Furthermore, the proposed RFM frameworkexhibits robust sim-to-real transfer capabilities and facilitates posetracking extensions with minimal architectural adaptation.
SingRef6D: Monocular Novel Object Pose Estimation with a Single RGBReference
- 分类：任意物体方法
- 论文：https://neurips.cc/virtual/2025/poster/115633
- 代码：代码未开源
- 摘要：Recent 6D pose estimation methods demonstrate notableperformance but still face some practical limitations. For instance,many of them rely heavily on sensor depth, which may fail withchallenging surface conditions, such as transparent or highly reflectivematerials. In the meantime, RGB-based solutions provide less robustmatching performance in low-light and texture-less scenes due to thelack of geometry information. Motivated by these, we propose SingRef6D,a lightweight pipeline requiring only a single RGB image as a reference,eliminating the need for costly depth sensors, multi-view imageacquisition, or training view synthesis models and neural fields. Thisenables SingRef6D to remain robust and capable even underresource-limited settings where depth or dense templates areunavailable. Our framework incorporates two key innovations. First, wepropose a token-scaler-based fine-tuning mechanism with a noveloptimization loss on top of Depth-Anything v2 to enhance its ability topredict accurate depth, even for challenging surfaces. Our results showa 14.41% improvement in (δ1.05) on REAL275 depth prediction compared toDepth-Anything v2 (with fine-tuned head). Second, benefiting from depthavailability, we introduce a depth-aware matching process thateffectively integrates spatial relationships within LoFTR, enabling oursystem to handle matching for challenging materials and lightingconditions. Evaluations of pose estimation on the REAL275, ClearPose,and Toyota-Light datasets show that our approach surpassesstate-of-the-art methods, achieving a 6.1% improvement in averagerecall.

ICLR

6D Object Pose Tracking in Internet Videos for Robotic Manipulation
- 分类：任意物体跟踪方法
- 论文：https://openreview.net/pdf?id=1CIUkpoata
- 代码：https://github.com/ponimatkin/freepose
- 摘要：We seek to extract a temporally consistent 6D pose trajectoryof a manipulated object from an Internet instructional video. This is achallenging set-up for current 6D pose estimation methods due touncontrolled capturing conditions, subtle but dynamic object motions,and the fact that the exact mesh of the manipulated object is not known.To address these challenges, we present the following contributions.First, we develop a new method that estimates the 6D pose of any objectin the input image without prior knowledge of the object itself. Themethod proceeds by (i) retrieving a CAD model similar to the depictedobject from a large-scale model database, (ii) 6D aligning the retrievedCAD model with the input image, and (iii) grounding the absolute scaleof the object with respect to the scene. Second, we extract smooth 6Dobject trajectories from Internet videos by carefully tracking thedetected objects across video frames. The extracted object trajectoriesare then retargeted via trajectory optimization into the configurationspace of a robotic manipulator. Third, we thoroughly evaluate and ablateour 6D pose estimation method on YCB-V and HOPE-Video datasets as wellas a new dataset of instructional videos manually annotated withapproximate 6D object trajectories. We demonstrate significantimprovements over existing state-of-the-art RGB 6D pose estimationmethods. Finally, we show that the 6D object motion estimated fromInternet videos can be transferred to a 7-axis robotic manipulator bothin a virtual simulator as well as in a real world set-up. We alsosuccessfully apply our method to egocentric videos taken from theEPIC-KITCHENS dataset, demonstrating potential for Embodied AIapplications.
Learning Shape-Independent Transformation via SphericalRepresentations for Category-Level Object Pose Estimation
- 分类：类别级方法
- 论文：https://openreview.net/pdf?id=D4xztKoz0Y
- 代码：代码未开源
- 摘要：Category-level object pose estimation aims to determine thepose and size of novel objects in specific categories. Existingcorrespondence-based approaches typically adopt point-basedrepresentations to establish the correspondences between primitiveobserved points and normalized object coordinates. However, due to theinherent shape-dependence of canonical coordinates, these methods sufferfrom semantic incoherence across diverse object shapes. To resolve thisissue, we innovatively leverage the sphere as a shared proxy shape ofobjects to learn shape-independent transformation via sphericalrepresentations. Based on this insight, we introduce a novelarchitecture called SpherePose, which yields precise correspondenceprediction through three core designs. Firstly, We endow the point-wisefeature extraction with SO(3)-invariance, which facilitates robustmapping between camera coordinate space and object coordinate spaceregardless of rotation transformation. Secondly, the spherical attentionmechanism is designed to propagate and integrate features amongspherical anchors from a comprehensive perspective, thus mitigating theinterference of noise and incomplete point cloud. Lastly, a hyperboliccorrespondence loss function is designed to distinguish subtledistinctions, which can promote the precision of correspondenceprediction. Experimental results on CAMERA25, REAL275 and HouseCat6Dbenchmarks demonstrate the superior performance of our method, verifyingthe effectiveness of spherical representations and architecturalinnovations.

ICML

ICCV

6DOPE-GS: Online 6D Object Pose Estimation using Gaussian Splatting
- 分类：任意物体方法
- 论文：https://openaccess.thecvf.com/content/ICCV2025/papers/Jin_6DOPE-GS_Online_6D_Object_Pose_Estimation_using_Gaussian_Splatting_ICCV_2025_paper.pdf
- 代码：代码未开源
- 摘要：Efficient and accurate object pose estimation is an essentialcomponent for modern vision systems in many applications such asAugmented Reality, autonomous driving, and robotics. While research inmodel-based 6D object pose estimation has delivered promising results,model-free methods are hindered by the high computational load inrendering and inferring consistent poses of arbitrary objects in a liveRGB-D video stream. To address this issue, we present 6DOPE-GS, a novelmethod for online 6D object pose estimation & tracking with a singleRGB-D camera by effectively leveraging advances in Gaussian Splatting.Thanks to the fast differentiable rendering capabilities of GaussianSplatting, our method can simultaneously optimize for 6D object posesand 3D object reconstruction. To achieve the necessary efficiency andaccuracy for live tracking, our method uses incremental 2D GaussianSplatting with an intelligent dynamic keyframe selection procedure toachieve high spatial object coverage and prevent erroneous pose updates.We also propose an opacity statistic-based pruning mechanism foradaptive Gaussian density control, to ensure training stability andefficiency. We evaluate our method on the HO3D and YCBInEOAT datasetsand show that 6DOPE-GS matches the performance of state-of-the-artbaselines for model-free simultaneous 6D pose tracking andreconstruction while providing a 5x speedup. We also demonstrate themethod’s suitability for live, dynamic object tracking andreconstruction in a real-world setting.
BoxDreamer: Dreaming Box Corners for Generalizable Object PoseEstimation
- 分类：任意物体方法
- 论文：https://openaccess.thecvf.com/content/ICCV2025/papers/Yu_BoxDreamer_Dreaming_Box_Corners_for_Generalizable_Object_Pose_Estimation_ICCV_2025_paper.pdf
- 代码：https://github.com/zju3dv/BoxDreamer
- 摘要：This paper presents a generalizable RGB-based approach forobject pose estimation, specifically designed to address challenges insparse-view settings. While existing methods can estimate the poses ofunseen objects, their generalization ability remains limited inscenarios involving occlusions and sparse reference views, restrictingtheir real-world applicability. To overcome these limitations, weintroduce corner points of the object bounding box as an intermediaterepresentation of the object pose. The 3D object corners can be reliablyrecovered from sparse input views, while the 2D corner points in thetarget view are estimated through a novel reference-based pointsynthesizer, which works well even in scenarios involving occlusions. Asobject semantic points, object corners naturally establish 2D-3Dcorrespondences for object pose estimation with a PnP algorithm.Extensive experiments on the YCB-Video and Occluded-LINEMOD datasetsshow that our approach outperforms state-of-the-art methods,highlighting the effectiveness of the proposed representation andsignificantly enhancing the generalization capabilities of object poseestimation, which is crucial for real-world applications.
CleanPose: Category-Level Object Pose Estimation via Causal Learningand Knowledge Distillation
- 分类：类别级方法
- 论文：https://openaccess.thecvf.com/content/ICCV2025/papers/Lin_CleanPose_Category-Level_Object_Pose_Estimation_via_Causal_Learning_and_Knowledge_ICCV_2025_paper.pdf
- 代码：https://github.com/chrislin0621/CleanPose
- 摘要：In the effort to achieve robust and generalizablecategory-level object pose estimation, recent methods primarily focus onlearning fundamental representations from data. However, the inherentbiases within the data are often overlooked: the repeated trainingsamples and similar environments may mislead the models to over-rely onspecific patterns, hindering models’ performance on novel instances. Inthis paper, we present CleanPose, a novel method that mitigates the databiases to enhance category-level pose estimation by integrating causallearning and knowledge distillation. By incorporating key causalvariables (structural information and hidden confounders) into causalmodeling, we propose the causal inference module based on front-dooradjustment, which promotes unbiased estimation by reducing potentialspurious correlations. Additionally, to further confront the data biasat the feature level, we devise a residual-based knowledge distillationapproach to transfer unbiased semantic knowledge from 3D foundationmodel, providing comprehensive causal supervision. Extensive experimentsacross multiple benchmarks (REAL275, CAMERA25 and HouseCat6D) highlightthe superiority of proposed CleanPose over state-of-the-art methods.Code will be available at https://github.com/chrislin0621/CleanPose.
Deterministic Object Pose Confidence Region Estimation
- 分类：实例级方法
- 论文：https://arxiv.org/pdf/2506.22720
- 代码：代码未开源
- 摘要：6D pose confidence region estimation has emerged as a criticaldirection, aiming to perform uncertainty quantification for assessingthe reliability of estimated poses. However, current sampling-basedapproach suffers from critical limitations that severely impede theirpractical deployment: 1) the sampling speed significantly decreases asthe number of samples increases. 2) the derived confidence regions areoften excessively large. To address these challenges, we propose adeterministic and efficient method for estimating pose confidenceregions. Our approach uses inductive conformal prediction to calibratethe deterministically regressed Gaussian keypoint distributions into 2Dkeypoint confidence regions. We then leverage the implicit functiontheorem to propagate these keypoint confidence regions directly into 6Dpose confidence regions. This method avoids the inefficiency andinflated region sizes associated with sampling and ensembling. Itprovides compact confidence regions that cover the ground-truth poseswith a user-defined confidence level. Experimental results on theLineMOD Occlusion and SPEED datasets show that our method achieveshigher pose estimation accuracy with reduced computational time. For thesame coverage rate, our method yields significantly smaller confidenceregion volumes, reducing them by up to 99.9% for rotations and 99.8% fortranslations. The code will be available soon.
Environment-Agnostic Pose: Generating Environment-independent ObjectRepresentations for 6D Pose Estimation
- 分类：实例级方法
- 论文：https://openaccess.thecvf.com/content/ICCV2025/papers/Zhang_Environment-Agnostic_Pose_Generating_Environment-independent_Object_Representations_for_6D_Pose_Estimation_ICCV_2025_paper.pdf
- 代码：https://github.com/acmff22/EA6D
- 摘要：This paper introduces EA6D, a novel diffusion-based frameworkfor 6D pose estimation that operates effectively in any environment.Traditional pose estimation methods struggle with the variability andcomplexity of real-world scenarios, often leading to overfitting oncontrolled datasets and poor generalization to new scenes. To addressthese challenges, we propose a generative pose estimation paradigm thatgenerates environment-independent object representations for poseestimation, which are robust to environmental variations such asillumination, occlusion, and background clutter. Specifically, wepropose the novel Environment Decoupling Diffusion Model (EDDM) whichseparates object representations from environmental factors whileenabling efficient few-step sampling by leveraging input image priorsinstead of pure noise initialization. We validate our approach on fourstandard benchmarks and a self-made dataset DiverseScenes. The resultsdemonstrate that EA6D, trained using only synthetic data, can outperformthe state-of-the-art methods with both synthetic and realistic data. Inparticular, for fair comparisons with synthetic data, we can exceed theprevious SOTA by 18.1% and 33.5% on Linemod and Linemod-Occludeddatasets respectively. Project page: https://github.com/acmff22/EA6D.
HccePose(BF): Predicting Front & Back Surfaces to ConstructUltra-Dense 2D-3D Correspondences for Pose Estimation
- 分类：实例级方法
- 论文：https://openaccess.thecvf.com/content/ICCV2025/papers/Wang_HccePoseBF_Predicting_Front__Back_Surfaces_to_Construct_Ultra-Dense_2D-3D_ICCV_2025_paper.pdf
- 代码：https://github.com/WangYuLin-SEU/HCCEPose
- 摘要：In pose estimation for seen objects, a prevalent pipelineinvolves using neural networks to predict dense 3D coordinates of theobject surface on 2D images, which are then used to establish dense2D-3D correspondences. However, current methods primarily focus on moreefficient encoding techniques to improve the precision of predicted 3Dcoordinates on the object’s front surface, overlooking the potentialbenefits of incorporating the back surface and interior of the object.To better utilize the full surface and interior of the object, thisstudy predicts 3D coordinates of both the object’s front and backsurfaces and densely samples 3D coordinates between them. This processcreates ultra-dense 2D-3D correspondences, effectively enhancing poseestimation accuracy based on the Perspective-n-Point (PnP) algorithm.Additionally, we propose Hierarchical Continuous Coordinate Encoding(HCCE) to provide a more accurate and efficient representation of frontand back surface coordinates. Experimental results show that, comparedto existing state-of-the-art (SOTA) methods on the BOP website, theproposed approach outperforms across seven classic BOP core datasets.Code is available at https://github.com/WangYuLin-SEU/HCCEPose.
Joint Learning of Pose Regression and Denoising Diffusion with ScoreScaling Sampling for Category-level 6D Pose Estimation
- 分类：类别级方法
- 论文：https://openaccess.thecvf.com/content/ICCV2025/papers/Lee_Joint_Learning_of_Pose_Regression_and_Denoising_Diffusion_with_Score_ICCV_2025_paper.pdf
- 代码：代码未开源
- 摘要：Latest diffusion models have shown promising results incategory-level 6D object pose estimation by modeling the conditionalpose distribution with depth image input. The existing methods, however,suffer from slow convergence during training, learning its encoder withthe diffusion denoising network in end-to-end fashion, and require anadditional network that evaluates sampled pose hypotheses to filter outlow-quality pose candidates. In this paper, we propose a novel pipelinethat tackles these limitations by two key components. First, theproposed method pre-trains the encoder with the direct pose regressionhead, and jointly learns the networks via the regression head and thedenoising diffusion head, significantly accelerating trainingconvergence while achieving higher accuracy. Second, sampling guidancevia time-dependent score scaling is proposed s.t. theexploration-exploitation trade-off is effectively taken, eliminating theneed for the additional evaluation network. The sampling guidancemaintains multimodal characteristics of symmetric objects at earlydenoising steps while ensuring high-quality pose generation at finalsteps. Extensive experiments on multiple benchmarks including REAL275,HouseCat6D, and ROPE, demonstrate that the proposed method, simple yeteffective, achieves state-of-the-art accuracies even with single-poseinference, while being more efficient in both training andinference.
MixRI: Mixing Features of Reference Images for Novel Object PoseEstimation
- 分类：任意物体方法
- 论文：https://openaccess.thecvf.com/content/ICCV2025/papers/Liu_MixRI_Mixing_Features_of_Reference_Images_for_Novel_Object_Pose_ICCV_2025_paper.pdf
- 代码：代码未开源
- 摘要：We present MixRI, a lightweight network that solves theCAD-based novel object pose estimation problem in RGB images. It can beinstantly applied to a novel object at test time without finetuning. Wedesign our network to meet the demands of real-world applications,emphasizing reduced memory requirements and fast inference time. Unlikeexisting works that utilize many reference images and have large networkparameters, we directly match points based on the multi-view informationbetween the query and reference images with a lightweight network.Thanks to our reference image fusion strategy, we significantly decreasethe number of reference images, thus decreasing the time needed toprocess these images and the memory required to store them. Furthermore,with our lightweight network, our method requires less inference time.Though with fewer reference images, experiments on seven core datasetsin the BOP challenge show that our method achieves comparable resultswith other methods that require more reference images and larger networkparameters.
RayPose: Ray Bundling Diffusion for Template Views in Unseen 6DObject Pose Estimation
- 分类：任意物体方法
- 论文：https://openaccess.thecvf.com/content/ICCV2025/papers/Huang_RayPose_Ray_Bundling_Diffusion_for_Template_Views_in_Unseen_6D_ICCV_2025_paper.pdf
- 代码：代码未开源
- 摘要：Typical template-based object pose pipelines estimate the poseby retrieving the closest matching template and aligning it with theobserved image. However, failure to retrieve the correct template oftenleads to inaccurate pose predictions. To address this, we reformulatetemplate-based object pose estimation as a ray alignment problem, wherethe viewing directions from multiple posed template images are learnedto align with a non-posed query image. Inspired by recent progress indiffusion-based camera pose estimation, we embed this formulation into adiffusion transformer architecture that aligns a query image with a setof posed templates. We reparameterize object rotation usingobject-centered camera rays and model object translation by extendingscale-invariant translation estimation to dense translation offsets. Ourmodel leverages geometric priors from the templates to guide accuratequery pose inference. A coarse-to-fine training strategy based onnarrowed template sampling improves performance without modifying thenetwork architecture. Extensive experiments across multiple benchmarkdatasets show competitive results of our method compared tostate-of-the-art approaches in unseen object pose estimation.
Unified Category-Level Object Detection and Pose Estimation from RGBImages using 3D Prototypes
- 分类：类别级方法
- 论文：https://openaccess.thecvf.com/content/ICCV2025/papers/Fischer_Unified_Category-Level_Object_Detection_and_Pose_Estimation_from_RGB_Images_ICCV_2025_paper.pdf
- 代码：https://github.com/Fischer-Tom/unified-detection-and-pose-estimation
- 摘要：Recognizing objects in images is a fundamental problem incomputer vision. Although detecting objects in 2D images is common, manyapplications require determining their pose in 3D space. Traditionalcategory-level methods rely on RGB-D inputs, which may not always beavailable, or employ two-stage approaches that use separate models andrepresentations for detection and pose estimation. For the first time,we introduce a unified model that integrates detection and poseestimation into a single framework for RGB images by leveraging neuralmesh models with learned features and multi-model RANSAC. Our approachachieves state-of-the-art results for RGB category-level pose estimationon REAL275, improving on the current state-of-the-art by 22.9% averagedacross all scale-agnostic metrics. Finally, we demonstrate that ourunified method exhibits greater robustness compared to single-stagebaselines. Our code and models are available atgithub.com/Fischer-Tom/unified-detection-and-pose-estimation.

AAAI

Free-Moving Object Reconstruction and Pose Estimation with VirtualCamera
- 分类：任意物体方法
- 论文：https://ojs.aaai.org/index.php/AAAI/article/view/32736
- 代码：https://github.com/HaixinShi/fmov_pose
- 摘要：We propose an approach for reconstructing free-moving objectfrom a monocular RGB video. Most existing methods either assume sceneprior, hand pose prior, object category pose prior, or rely on localoptimization with multiple sequence segments. We propose a method thatallows free interaction with the object in front of a moving camerawithout relying on any prior, and optimizes the sequence globallywithout any segments. We progressively optimize the object shape andpose simultaneously based on an implicit neural representation. A keyaspect of our method is a virtual camera system that reduces the searchspace of the optimization significantly. We evaluate our method on thestandard HO3D dataset and a collection of egocentric RGB sequencescaptured with a head-mounted device. We demonstrate that our approachoutperforms most methods significantly, and is on par with recenttechniques that assume prior information.
KeyPose: Category-Level 6D Object Pose Estimation with Self-AdaptiveKeypoints
- 分类：类别级方法
- 论文：https://ojs.aaai.org/index.php/AAAI/article/view/33046
- 代码：https://github.com/BIT-robot-group/KeyPose
- 摘要：Category-level object pose estimation is an important task incomputer vision. Some prior methods based on assumptions often strugglewith drastic changes in object appearance. To address this challenge, wepropose a new method for object pose estimation based on object-adaptivekeypoints. In this paper, we first introduce a transformer-basedkeypoint prediction method for adaptive forecasting of point cloudkeypoints. This method calculates the similarity between keypointfeatures and point cloud features, allowing keypoints to representobject geometry more effectively. Furthermore, to enhance the geometricfeature construction of keypoints, we propose a graph-based keypointfeature aggregation method, which considers both the structuralrelationships between keypoints and the point cloud, strengthening thenetwork’s understanding of geometric structures. At this stage,keypoints remain at the geometric spatial level of the object and havenot been predicted in NOCS. To improve the accuracy of keypointprediction in NOCS, we design a NOCS voxelization method that dividesNOCS into multiple voxels and accurately predicts NOCS keypoints withinthese voxels. Experimental results on multiple benchmark datasetsdemonstrate that our proposed KeyPose method outperforms all existingmethods, achieving over 20% improvement in pose accuracy on somecritical datasets.
R²-Art: Category-Level Articulation Pose Estimation from Single RGBImage via Cascade Render Strategy
- 分类：类别级铰接部件方法
- 论文：https://ojs.aaai.org/index.php/AAAI/article/view/33083
- 代码：代码未开源
- 摘要：Human life is filled with articulated objects. Previous worksfor estimating the pose of category-level articulated objects rely oncostly 3D point clouds or RGB-D images. In this paper, our goal is toestimate category-level articulation poses from a single RGB image,where we propose R²-Art, a novel category-level Articulation poseestimation framework from a single RGB image and a cascade Renderstrategy. Given an RGB image as input, R²-Art estimates per-part 6D posefor the articulation. Specifically, we design parallel regressionbranches tailored to generate camera-to-root translation and rotation.Using the predicted joint states, we perform PC prior transformation anddeformation with a joint-centric modeling approach. For furtherrefinement, a cascade render strategy is proposed for projecting the 3Ddeformed prior onto the 2D mask. Extensive experiments are provided tovalidate our R²-Art on various datasets ranging from synthetic datasetsto real-world scenarios, demonstrating the superior performance androbustness of the R²-Art. We believe that this work has the potential tobe applied in many fields including robotics, embodied intelligence, andaugmented reality.
Universal Features Guided Zero-Shot Category-Level Object PoseEstimation
- 分类：类别级方法
- 论文：https://ojs.aaai.org/index.php/AAAI/article/view/32713
- 代码：代码未开源
- 摘要：Object pose estimation, crucial in computer vision androbotics applications, faces challenges with the diversity of unseencategories. We propose a zero-shot method to achieve category-level6-DOF object pose estimation, which exploits both 2D and 3D universalfeatures of input RGB-D image to establish semantic similarity-basedcorrespondences and can be extended to unseen categories withoutadditional model fine-tuning. Our method begins with combining efficient2D universal features to find sparse correspondences betweenintra-category objects and gets initial coarse pose. To handle thecorrespondence degradation of 2D universal features if the pose deviatesmuch from the target pose, we use an iterative strategy to optimize thepose. Subsequently, to resolve pose ambiguities due to shape differencesbetween intra-category objects, the coarse pose is refined by optimizingwith dense alignment constraint of 3D universal features. Our methodoutperforms previous methods on the REAL275 and Wild6D benchmarks forunseen categories.

SIGGRAPH

ACM MM

CLIP-6D: Empowering CLIP as a Zero-Shot 6D Pose Estimator ThroughGeneralizable Object-Specific Representations
- 分类：任意物体方法
- 论文：https://dl.acm.org/doi/10.1145/3746027.3754497
- 代码：代码未开源
- 摘要：The open-vocabulary paradigm enhances 6D object poseestimation by leveraging language cues to transfer learnedrepresentations from seen to unseen objects, yet its performance suffersfrom a misalignment between vision-language representations and the 6Dpose space. This challenge is compounded by the intrinsic lack of poseawareness in models like CLIP (Contrastive Language-Image Pre-Training).To address this limitation, we introduce CLIP-6D, which equips CLIP withthe ability to estimate poses through the learning of generalizablerepresentations specific to objects. CLIP-6D consists of three keycomponents: (1) an innovative self-alignment strategy that enables CLIPto derive geometric representations from RGB images by leveraging itsinherent feature extractor; (2) a multiplex representation interactivelearning method efficiently bridges the heterogeneous representations ofobject category priors, geometry, and spatial correlations; (3) alightweight adapter using knowledge distillation improves CLIP’s captureof detailed semantic representations. Experiments show that CLIP-6Dachieves an improvement of 10.8% and 16.4% in the metric 5° 2 cm and 10°5 cm for zero-shot generalization and achieves a speedup of 5.4 FPS ininference over current state-of-the-art methods. The source code andmodels are available at https://github.com/whoawong/CLIP-6D.git.

CCF B

ECCV

IJCAI

Pre-defined Keypoints Promote Category-level Articulation PoseEstimation via Multi-Modal Alignment
- 分类：类别级铰接部件方法
- 论文：https://www.ijcai.org/proceedings/2025/0237.pdf
- 代码：代码未开源
- 摘要：Articulations are essential in everyday interactions, yettraditional RGB-based pose estimation methods often struggle with issuessuch as lighting variations and shadows. To overcome these challenges,we propose a novel Pre-defined keypoint based framework forcategory-level articulation pose estimation via multi-modal AliGnMent,coined PAGE. Specifically, we first propose a customized keypointestimation method, aiming to avoid the divergent distance patternbetween heuristically generated keypoints and visible points. Inaddition, to reduce the mutual information redundancy between pointclouds and RGB images, we design the geometry-color alignment, whichfuses the features after aligning two modalities. This is followed bydecoding the radius for each visible point, and applying our proposalintegration scoring strategy to predict keypoints. Ultimately, theframework outputs the per-part 6D pose of the articulation. We conductextensive experiments to evaluate PAGE across a variety of datasets,from synthetic to real-world scenarios, demonstrating its robustness andsuperior performance.
SeqPose: An End-to-End Framework to Unify Single-frame andVideo-based RGB Category-Level Pose Estimation
- 分类：类别级方法
- 论文：https://www.ijcai.org/proceedings/2025/0137.pdf
- 代码：代码未开源
- 摘要：Category-level object pose estimation is a long-standing andfundamental task crucial for augmented reality and robotic manipulationapplications. Existing RGB-based approaches struggle with multi-stagesettings and heavily rely on off-the-shelf techniques, such as objectdetectors, depth estimators, non-differentiable NOCS shape alignment,etc. Extra dependencies lead to the accumulation of errors andcomplicate the whole pipeline, limiting the deployment of theseapproaches in practical applications. This paper streamlined anend-to-end framework unifying the single-frame and video-basedcategory-level pose estimation. Specifically, instead of explicitlyintroducing extra dependencies, the DINOv2 encoder and depth decoder, asrobust semantic and geometric prior extractors, are leveraged to produceintra-frame hierarchical semantic and geometric features. Aspatial-temporal sparse query network is developed to model the implicitcorrespondence and inter-frame correlations between a set of implicit 3Dquery anchors and intra-frame features. Finally, a pose prediction headis employed using the bipartite matching algorithm. Experimental resultsdemonstrate that our model achieves state-of-the-art performancecompared with RGB-based categorical pose estimation methods on theREAL275 and CAMERA25 datasets. Our code is available at https://andrewchiyz.github.io/vision.3dv.seqpose/.

ICRA

A Unified End-to-End Network for Category-Level and Instance-LevelObject Pose Estimation from RGB Images
- 分类：实例级&类别级方法
- 论文：https://ieeexplore.ieee.org/document/11128247
- 代码：代码未开源
- 摘要：Accurately estimating the 6-DoF pose of objects is afundamental challenge in computer vision and robotics. Whilecategory-level pose estimation based on RGBD data has achieved goodperformance in recent years, estimating poses solely from RGB imagesremains a significant challenge. Existing RGB-based category-levelmethods primarily focus on recovering object point clouds from RGBimages, and pose prediction is not performed end-to-end by a network.This paper presents a Category-level and Instance-level Pose EstimationNetwork (CIPE), which models pose estimation as a set prediction problemand enables direct pose regression from RGB images. To further enhancethe network’s ability to learn object poses, first, a novel learnablerotation representation that redefines rotation learning withinEuclidean space is introduced to facilitate rotation regression.Additionally, we propose a prior-query fusion strategy that utilizes apre-trained point cloud feature extraction network to integratecategorical object features with bounding boxes, thereby improving theincorporation of category information. Experimental results demonstratethat CIPE significantly outperforms existing RGB-based methods on bothcategory-level and instance-level datasets. The code is available at https://github.com/jialeren/CIPE.
MonoDiff9D: Monocular Category-Level 9D Object Pose Estimation viaDiffusion Model
- 分类：类别级方法
- 论文：https://ieeexplore.ieee.org/document/11127837
- 代码：https://github.com/CNJianLiu/MonoDiff9D
- 摘要：Object pose estimation is a core means for robots tounderstand and interact with their environment. For this task, monocularcategory-level methods are attractive as they require only a single RGBcamera. However, current methods rely on shape priors or CAD models ofthe intra-class known objects. We propose a diffusion-based monocularcategory-level 9D object pose generation method, MonoDiff9D. Ourmotivation is to leverage the probabilistic nature of diffusion modelsto alleviate the need for shape priors, CAD models, or depth sensors forintra-class unknown object pose estimation. We first estimate coarsedepth via DINOv2 from the monocular image in a zero-shot manner andconvert it into a point cloud. We then fuse the global features of thepoint cloud with the input image and use the fused features along withthe encoded time step to condition MonoDiff9D. Finally, we design atransformer-based denoiser to recover the object pose from Gaussiannoise. Extensive experiments on two popular benchmark datasets show thatMonoDiff9D achieves state-of-the-art monocular category-level 9D objectpose estimation accuracy without the need for shape priors or CAD modelsat any stage. Our code will be made public at https://github.com/CNJianLiu/MonoDiff9D.
Occlusion-Aware 6D Pose Estimation with Depth-Guided Graph Encodingand Cross-Semantic Fusion for Robotic Grasping
- 分类：实例级方法
- 论文：https://ieeexplore.ieee.org/document/11128248
- 代码：代码未开源
- 摘要：Reliable 6D pose estimation is crucial for robotic tasks butpresents significant challenges in environments with occlusion. Recentapproaches tend to directly predict pose parameters of object with deepneural networks, lacking the modeling ability of non-adjacent andcomplex relationships of surface points in occluded scenarios. To solvethis problem, we propose a novel occlusion-aware 6D pose estimationframework, which uses depth-guided graph neural network (GNN) to modelpotential relationships from RGBD input. Two semantic information, whichare mask and binary code of object, are adaptively fused to extract2D-3D correspondence related features in an effective manner. Bothenhanced graph features and fused semantic information contribute to theperformance improvement of pose estimation with occlusion. Extensiveexperiments indicate that our approach outperforms comparative methodsby 1.2% and 1.9% on LMO and YCBV datasets (up to 30% for certainobjects) and its validity is also verified under real-world poseestimation test.
OmniShape: Zero-Shot Multi-Hypothesis Object Pose and ShapeEstimation in the Real World
- 分类：任意物体方法
- 论文：https://ieeexplore.ieee.org/document/11128589
- 代码：代码未开源
- 摘要：We would like to estimate the pose and full shape of an objectfrom a single observation, without assuming known 3D model or category.In this work, we propose OmniShape, the first method of its kind toenable probabilistic pose and shape estimation. OmniShape is based onthe key insight that shape completion can be decoupled into twomulti-modal distributions: one capturing how measurements project into anormalized object reference frame defined by the dataset and the othermodelling a prior over object geometries represented as triplanar neuralfields. By training separate conditional diffusion models for these twodistributions, we enable sampling multiple hypotheses from the jointpose and shape distribution. OmniShape demonstrates compellingperformance on challenging real world datasets. Project website: https://tri-ml.github.io/omnishape.
You Only Estimate Once: Unified, One-stage, Real-Time Category-LevelArticulated Object 6D Pose Estimation for Robotic Grasping
- 分类：类别级铰接部件方法
- 论文：https://ieeexplore.ieee.org/document/11128513
- 代码：代码未开源
- 摘要：This paper addresses the problem of category-level poseestimation for articulated objects in robotic manipulation tasks. Recentworks have shown promising results in estimating part pose and size atthe category level. However, these approaches primarily follow a complexmulti-stage pipeline that first segments part instances in the pointcloud and then estimates the Normalized Part Coordinate Space (NPCS)representation for 6D poses. These approaches suffer from highcomputational costs and low performance in real-time robotic tasks. Toaddress these limitations, we propose YOEO, a single-stage method thatsimultaneously outputs instance segmentation and NPCS representations inan end-to-end manner. We use a unified network to generate point-wisesemantic labels and centroid offsets, allowing points from the same partinstance to vote for the same centroid. We further utilize a clusteringalgorithm to distinguish points based on their estimated centroiddistances. Finally, we first separate the NPCS region of each instance.Then, we align the separated regions with the real point cloud torecover the final pose and size. Experimental results on the GAPartdataset demonstrate the pose estimation capabilities of our proposedsingle-shot method. We also deploy our synthetically-trained model in areal-world setting, providing real-time visual feedback at 200Hz,enabling a physical Kinova robot to interact with unseen articulatedobjects. This showcases the utility and effectiveness of ourmethod.

ICME

Diff-Art: Category-level Articulation Pose Estimation viaConditional Diffusion
- 分类：类别级铰接部件方法
- 论文：https://ieeexplore.ieee.org/document/11209596
- 代码：代码未开源
- 摘要：Articulated objects are prevalent in people’s daily lives, yettheir diverse motion structures pose significant challenges forcategory-level pose estimation. To address this, in this work, weintroduce Diff-Art, a method tailored for category-level articulationpose estimation using conditional diffusion. Given a partial point cloudas input, Diff-Art predicts the per-part 6D pose of the articulatedobject. Our approach incorporates a novel modeling strategy thatexploits the unique kinematic constraints of articulated objects,effectively handling self-occlusion scenarios. Furthermore, we propose are-scoring strategy to enhance the accuracy of 6D pose estimation withinthe conditional diffusion framework. Extensive experiments validate theeffectiveness of Diff-Art, demonstrating its strong performance on bothsynthetic datasets and its ability to generalize to real-worldscenarios. We believe this work holds significant potential forapplications in embodied AI and robotics.
GASEM: Boosting Generalized and Actionable Parts Segmentation andPose Estimation via Object Motion Perception
- 分类：类别级可操作部件方法
- 论文：https://ieeexplore.ieee.org/document/11208907
- 代码：https://github.com/Zirconium233/GASEM
- 摘要：Category-level object understanding has progressed, butgeneralized part perception remains underexplored. This paper introducesGASEM, a framework for Generalizable and Actionable Parts GAPartSegmentation and pose Estimation via object Motion perception. GASEMutilizes point-wise motion data from observed point clouds andcross-perspective alignment to learn object motion using a scene flowmodel. It features a segmentation proposal architecture for GAPartsegmentation and an Normalized Object Coordinate Space(NPCS) branch forpose estimation. Additionally, a reinforcement learning agent is trainedfor robust GAPart manipulation in both simulations and real-worldenvironments. Experiments on the GAPartNet dataset show GASEMoutperforms state-of-the-art methods. This work promises advancements inembodied intelligence applications like robot-object interaction andgeneralizable manipulation. Codes are available at https://github.com/Zirconium233/GASEM.
Multi-mode Bidirectional Feature Fusion and Domain-consistencyRefinement for Real-time Monocular 6D Object Pose Estimation
- 分类：实例级方法
- 论文：https://ieeexplore.ieee.org/document/11208924
- 代码：代码未开源
- 摘要：Current pose estimation methods are confronted computationalperformance, domain gap in refinement and feature sharing among varioussemantic tasks. To address these challenges, we present a lightweightpose estimation framework with the following three novel contributions.Firstly, we propose a novel for real-time object pose estimation,composed of the multi-task network called DSPNet to exploit semanticrelations among different tasks and 3D Gaussian based pose refinement toaddress the domain gap problem. Secondly, the both networks learntemporal features of previous camera trajectory by a transformer basedmodule to further improve estimation robustness. Thirdly, we design abidirectional fusion module between consecutive decoding layers inDSPNet to achieve two-sided feature sharing including various semanticfeatures and spatio-temporal features. Experimental results on thepublic LineMod, LineMod-Occluded and YCB-Video datasets demonstrate thestate-of-the-art performance in comparison with other RGB based methodswhile keeping 37 FPS on an Nvidia GTX 3090 GPU.

ECAI

ISMAR

CCF C

AISTATS

IROS

6-DoF Object Tracking with Event-based Optical Flow and Frames
- 分类：实例级跟踪方法
- 论文：https://ieeexplore.ieee.org/document/11245865
- 代码：代码未开源
- 摘要：Tracking the position and orientation of objects in space(i.e., in 6-DoF) in real time is a fundamental problem in robotics forenvironment interaction. It becomes more challenging when objects moveat high-speed due to frame rate limitations in conventional cameras andmotion blur. Event cameras are characterized by high temporalresolution, low latency and high dynamic range, that can potentiallyovercome the impacts of motion blur. Traditional RGB cameras providerich visual information that is more suitable for the challenging taskof single-shot object pose estimation. In this work, we propose usingevent-based optical flow combined with an RGB based global object poseestimator for 6-DoF pose tracking of objects at high-speed, exploitingthe core advantages of both types of vision sensors. Specifically, wepropose an event-based optical flow algorithm for object motionmeasurement to implement an object 6-DoF velocity tracker. Byintegrating the tracked object 6-DoF velocity with low frequencyestimated pose from the global pose estimator, the method can track posewhen objects move at high-speed. The proposed algorithm is tested andvalidated on both synthetic and real world data, demonstrating itseffectiveness, especially in high-speed motion scenarios.
Distilling 3D distinctive local descriptors for 6D pose estimation
- 分类：任意物体方法
- 论文：https://ieeexplore.ieee.org/document/11247352
- 代码：https://github.com/tev-fbk/dGeDi/
- 摘要：Three-dimensional local descriptors are crucial for encodinggeometric surface properties, making them essential for various pointcloud understanding tasks. Among these descriptors, GeDi hasdemonstrated strong zero-shot 6D pose estimation capabilities butremains computationally impractical for real-world applications due toits expensive inference process. Can we retain GeDi’s effectivenesswhile significantly improving its efficiency? In this paper, we explorethis question by introducing a knowledge distillation framework thattrains an efficient student model to regress local descriptors from aGeDi teacher. Our key contributions include: an efficient large-scaletraining procedure that ensures robustness to occlusions and partialobservations while operating under compute and storage constraints, anda novel loss formulation that handles weak supervision fromnon-distinctive teacher descriptors. We validate our approach on fiveBOP Benchmark datasets and demonstrate a significant reduction ininference time while maintaining competitive performance with existingmethods, bringing zero-shot 6D pose estimation closer to real-timefeasibility. Project Website: https://tev-fbk.github.io/dGeDi/.
DynamicPose: Real-time and Robust 6D Object Pose Tracking forFast-Moving Cameras and Objects
- 分类：任意物体跟踪方法
- 论文：https://ieeexplore.ieee.org/document/11247403
- 代码：https://github.com/Robotics-STAR-Lab/DynamicPose
- 摘要：We present DynamicPose, a retraining-free 6D pose trackingframework that improves tracking robustness in fast-moving camera andobject scenarios. Previous work is mainly applicable to static orquasi-static scenes, and its performance significantly deteriorates whenboth the object and the camera move rapidly. To overcome thesechallenges, we propose three synergistic components: (1) Avisual-inertial odometry compensates for the shift in the Region ofInterest (ROI) caused by camera motion; (2) A depth-informed 2D trackercorrects ROI deviations caused by large object translation; (3) AVIO-guided Kalman filter predicts object rotation, generates multiplecandidate poses, and then obtains the final pose by hierarchicalrefinement. The 6D pose tracking results guide subsequent 2D trackingand Kalman filter updates, forming a closed-loop system that ensuresaccurate pose initialization and precise pose tracking. Simulation andreal-world experiments demonstrate the effectiveness of our method,achieving real-time and robust 6D pose tracking for fast-moving camerasand objects.
Generalizable and Actionable Part Detection and Manipulation withSAM-rectified Segmentation and Iterative Pose Refinement
- 分类：类别级可操作部件方法
- 论文：https://ieeexplore.ieee.org/document/11246094
- 代码：代码未开源
- 摘要：The ability to perform cross-category object perception andmanipulation is highly desirable in building intelligent robots. Onepromising approach is to define the concept of Generalizable andActionable Parts (GAParts), such as buttons and handles, on both seenand unseen object categories. However, the accurate cross-categoryperception of GAParts is still challenging due to the largeinter-category object shape variations. To address this issue, weintroduce SAMIR, a novel framework using SAM-rectified segmentation andIterative pose Refinement for GAPart detection and manipulation.Firstly, we introduce a Segment Anything (SAM) segmentation prior torectify the unconfident, fragmented GAPart instance proposals. Secondly,in addition to the zero-shot generalization of the SAM foundation model,we further finetune it with a lightweight adaptor model on our taskdataset. Finally, we propose an iterative pose refinement procedure thatimproves the accuracy of GAPart pose estimation. Our perceptionexperiments on GAPartNet dataset show that SAMIR consistentlyoutperforms the baseline method on instance segmentation and poseestimation tasks. Our manipulation experiments in Sapien simulatorillustrate that SAMIR leads to an improved manipulation success rate. Wealso deploy our method to a real robot for real-world manipulation. Ourcode and video are available at sites.google.com/view/samir-gapart.
Improving 6D Object Pose Estimation of Metallic Household andIndustry Objects
- 分类：实例级方法
- 论文：https://ieeexplore.ieee.org/document/11247314
- 代码：代码未开源
- 摘要：6D object pose estimation suffers from reduced accuracy whenapplied to metallic objects. We set out to improve the state-of-the-artby addressing challenges such as reflections and specular highlights inindustrial applications. Our novel BOP-compatible dataset, featuring adiverse set of metallic objects (cans, household, and industrial items)under various lighting and background conditions, provides additionalgeometric and visual cues. We demonstrate that these cues can beeffectively leveraged to enhance overall performance. To illustrate theusefulness of the additional features, we improve upon the GDRNPPalgorithm by introducing an additional keypoint prediction and materialestimator head in order to improve spatial scene understanding.Evaluations on the new dataset show improved accuracy for metallicobjects, supporting the hypothesis that additional geometric and visualcues can improve learning.
MK-Pose: Category-Level Object Pose Estimation via Multimodal-BasedKeypoint Learning
- 分类：类别级方法
- 论文：https://ieeexplore.ieee.org/document/11247325
- 代码：https://github.com/yangyifanYYF/MK-Pose
- 摘要：Category-level object pose estimation, which predicts the poseof objects within a known category without prior knowledge of individualinstances, is essential in applications like warehouse automation andmanufacturing. Existing methods relying on RGB images or point clouddata often struggle with object occlusion and generalization acrossdifferent instances and categories. This paper proposes amultimodal-based keypoint learning framework (MK-Pose) that integratesRGB images, point clouds, and category-level textual descriptions. Themodel uses a self-supervised keypoint detection module enhanced withattention-based query generation, soft heatmap matching and graph-basedrelational modeling. Additionally, a graph-enhanced feature fusionmodule is designed to integrate local geometric information and globalcontext. MK-Pose is evaluated on CAMERA25 and REAL275 dataset, and isfurther tested for cross-dataset capability on HouseCat6D dataset. Theresults demonstrate that MK-Pose outperforms existing state-of-the-artmethods in both IoU and average precision without shape priors. Codeswill be released at https://github.com/yangyifanYYF/MK-Pose.
Occlusion-Aware 6D Pose Estimation with Visual Observation GuidedDiffusion Model
- 分类：类别级方法
- 论文：https://ieeexplore.ieee.org/document/11247112
- 代码：代码未开源
- 摘要：Category-level 6D pose estimation in cluttered and occludedenvironments is a challenging task. Most existing methods rely ondeterministic point-based correspondences to estimate target poses,which cannot consider the uncertainty for occluded objects, and thusresult in inferior performance. In this paper, we propose a diffusionmodel guided by occlusion-aware observations to adaptively refine theobject poses in occluded and cluttered scenes. Specifically, we firstextract various 2D and 3D features from an RGB-D image to construct theconditions of diffusion model. In the reverse diffusion process, themodel is guided by implicit correspondences, perception distance, andocclusion relationships to refine the noisy pose sampled from a standardGaussian distribution. With several denoising steps, our method canproduce accurate results that are consistent with image observations inoccluded scenarios. The experimental results show that the proposedmethod can outperform baseline methods in major metrics in occlusionscenarios. Furthermore, our approach can also be applied in roboticgrasping and manipulation tasks through grasping experiments in acluttered enviroment on a physical UR5 robot.
OmniPose6D: Towards Short-Term Object Pose Tracking in DynamicScenes from Monocular RGB
- 分类：任意物体跟踪方法
- 论文：https://ieeexplore.ieee.org/document/11246949
- 代码：代码未开源
- 摘要：To address the challenge of short-term object pose tracking indynamic environments with monocular RGB input, we introduce alarge-scale synthetic dataset OmniPose6D, crafted to mirror thediversity of real-world conditions. We additionally present abenchmarking framework for a comprehensive comparison of pose trackingalgorithms. We propose a pipeline featuring an uncertainty-awarekeypoint refinement network, employing probabilistic modeling to refinepose estimation. Comparative evaluations demonstrate that our approachachieves performance superior to existing baselines on real datasets,underscoring the effectiveness of our synthetic dataset and refinementtechnique in enhancing tracking precision in dynamic contexts. Ourcontributions set a new precedent for the development and assessment ofobject pose tracking methodologies in complex scenes.
RAG-6DPose: Retrieval-Augmented 6D Pose Estimation via LeveragingCAD as Knowledge Base
- 分类：任意物体方法
- 论文：https://ieeexplore.ieee.org/document/11246449
- 代码：https://github.com/SresserS/RAG-6DPose-code
- 摘要：Accurate 6D pose estimation is key for robotic manipulation,enabling precise object localization for tasks like grasping. We presentRAG-6DPose, a retrieval-augmented approach that leverages 3D CAD modelsas a knowledge base by integrating both visual and geometric cues. OurRAG-6DPose roughly contains three stages: 1) Building a Multi-Modal CADKnowledge Base by extracting 2D visual features from multi-view CADrendered images and also attaching 3D points; 2) Retrieving relevant CADfeatures from the knowledge base based on the current query image viaour ReSPC module; and 3) Incorporating retrieved CAD information torefine pose predictions via retrieval-augmented decoding. Experimentalresults on standard benchmarks and real-world robotic tasks demonstratethe effectiveness and robustness of our approach, particularly inhandling occlusions and novel viewpoints. Supplementary material isavailable on our project website: https://sressers.github.io/RAG-6DPose.
RCGNet: RGB-based Category-Level 6D Object Pose Estimation withGeometric Guidance
- 分类：类别级方法
- 论文：https://ieeexplore.ieee.org/document/11247442
- 代码：代码未开源
- 摘要：While most current RGB-D-based category-level object poseestimation methods achieve strong performance, they face significantchallenges in scenes lacking depth information. In this paper, wepropose a novel category-level object pose estimation approach thatrelies solely on RGB images. This method enables accurate poseestimation in real-world scenarios without the need for depth data.Specifically, we design a transformer-based neural network forcategory-level object pose estimation, where the transformer is employedto predict and fuse the geometric features of the target object. Toensure that these predicted geometric features faithfully capture theobject’s geometry, we introduce a geometric feature-guided algorithm,which enhances the network’s ability to effectively represent theobject’s geometric information. Finally, we utilize the RANSAC-PnPalgorithm to compute the object’s pose, addressing the challengesassociated with variable object scales in pose estimation. Experimentalresults on benchmark datasets demonstrate that our approach is not onlyhighly efficient but also achieves superior accuracy compared toprevious RGB-based methods. These promising results offer a newperspective for advancing category-level object pose estimation usingRGB images.
RGBTrack: Fast, Robust Depth-Free 6D Pose Estimation and Tracking
- 分类：任意物体跟踪方法
- 论文：https://ieeexplore.ieee.org/document/11247381
- 代码：https://github.com/GreatenAnoymous/RGBTrack
- 摘要：We introduce a robust framework, RGBTrack, for real-time 6Dpose estimation and tracking that operates solely on RGB data, therebyeliminating the need for depth input for such dynamic and precise objectpose tracking tasks. Building on the FoundationPose architecture, wedevise a novel binary search strategy combined with a render-and-comparemechanism to efficiently infer depth and generate robust pose hypothesesfrom true-scale CAD models. To maintain stable tracking in dynamicscenarios, including rapid movements and occlusions, RGBTrack integratesstate-of-the-art 2D object tracking (XMem) with a Kalman filter and astate machine for proactive object pose recovery. In addition,RGBTrack’s scale recovery module dynamically adapts CAD models ofunknown scale using an initial depth estimate, enabling seamlessintegration with modern generative reconstruction techniques. Extensiveevaluations on benchmark datasets demonstrate that RGBTrack’s noveldepth-free approach achieves competitive accuracy and real-timeperformance, making it a promising practical solution candidate forapplication areas including robotics, augmented reality, and computervision. The source code for our implementation will be made publiclyavailable at https://github.com/GreatenAnoymous/RGBTrack.git.
Uncertainty-Aware Knowledge Distillation for Compact and Efficient6DoF Pose Estimation
- 分类：实例级方法
- 论文：https://ieeexplore.ieee.org/document/11245982
- 代码：代码未开源
- 摘要：Compact and efficient 6DoF object pose estimation is crucialin applications such as robotics, augmented reality, and spaceautonomous navigation systems, where lightweight models are critical forreal-time accurate performance. This paper introduces a noveluncertainty-aware end-to-end Knowledge Distillation (KD) frameworkfocused on keypoint-based 6DoF pose estimation. Keypoints predicted by alarge teacher model exhibit varying levels of uncertainty that can beexploited within the distillation process to enhance the accuracy of thestudent model while ensuring its compactness. To this end, we propose adistillation strategy that aligns the student and teacher predictions byadjusting the knowledge transfer based on the uncertainty associatedwith each teacher keypoint prediction. Additionally, the proposed KDleverages this uncertainty-aware alignment of keypoints to transfer theknowledge at key locations of their respective feature maps. Experimentson the widely-used LINEMOD benchmark demonstrate the effectiveness ofour method, achieving superior 6DoF object pose estimation withlightweight models compared to state-of-the-art approaches. Furthervalidation on the SPEED+ dataset for spacecraft pose estimationhighlights the robustness of our approach under diverse 6DoF poseestimation scenarios.
VAPO: Visibility-Aware Keypoint Localization for Efficient 6DoFObject Pose Estimation
- 分类：实例级方法
- 论文：https://ieeexplore.ieee.org/document/11247012
- 代码：https://github.com/RuyiLian/VAPO
- 摘要：Localizing predefined 3D keypoints in a 2D image is aneffective way to establish 3D-2D correspondences for 6DoF object poseestimation. However, unreliable localization results of invisiblekeypoints degrade the quality of correspondences. In this paper, weaddress this issue by localizing the important keypoints in terms ofvisibility. Since keypoint visibility information is currently missingin the dataset collection process, we propose an efficient way togenerate binary visibility labels from available object-levelannotations, for keypoints of both asymmetric objects and symmetricobjects. We further derive real-valued visibility-aware importance frombinary labels based on the PageRank algorithm. Taking advantage of theflexibility of our visibility-aware importance, we construct VAPO(Visibility-Aware POse estimator) by integrating the visibility-awareimportance with a state-of-the-art pose estimation algorithm, along withadditional positional encoding. VAPO can work in both CAD-based andCAD-free settings. Extensive experiments are conducted on popular poseestimation benchmarks including Linemod, Linemod-Occlusion, and YCB-V.The results show that, VAPO significantly improves both the keypointcorrespondences and final estimated poses, and clearly achievesstate-of-the-art performances.

ICPR

IJCNN

BMVC

3DV

GS-Pose: Generalizable Segmentation-Based 6D Object Pose EstimationWith 3D Gaussian Splatting
- 分类：任意物体方法
- 论文：https://openreview.net/pdf?id=8xHWWXZ9pq
- 代码：https://github.com/dingdingcai/GSPose
- 摘要：This paper introduces GS-Pose, a unified framework forlocalizing and estimating the 6D pose of novel objects. GS-Pose beginswith a set of posed RGB images of a previously unseen object and buildsthree distinct representations stored in a database. At inference,GS-Pose operates sequentially by locating the object in the input image,estimating its initial 6D pose using a retrieval approach, and refiningthe pose with a render-and-compare method. The key insight is theapplication of the appropriate object representation at each stage ofthe process. In particular, for the refinement step, we leverage 3DGaussian splatting, a novel differentiable rendering technique thatoffers high rendering speed and relatively low optimization time.Off-the-shelf toolchains and commodity hardware, such as mobile phones,can be used to capture new objects to be added to the database.Extensive evaluations on the LINEMOD and OnePose-LowTexture datasetsdemonstrate excellent performance, establishing the newstate-of-the-art. The source code is publicly available at https://github.com/dingdingcai/GSPose.

ICIP

Adapting Foundation Features via Cross-View Contrastive Learning forUnseen Object Pose Estimation
- 分类：任意物体方法
- 论文：https://ieeexplore.ieee.org/document/11084358
- 代码：代码未开源
- 摘要：Recent advances in visual foundation models, such as DINOv2,have shown impressive generalization capabilities, making them a strongbaseline for complex tasks like unseen object pose estimation. However,a significant domain gap between real images and rendered templates,combined with limited 3D awareness, hinders their effectiveness inprecise correspondence matching and pose estimation. In this paper, wepropose a new adaptation method that reduces this domain gap andenhances multi-view consistency through cross-view contrastive learning.Our approach aligns features from real and rendered images by usingsampled poses and 2D-2D correspondences while also leveraging advancedaugmentations and PCA-based loss computation to improve robustness andefficiency. When applied to the FoundPose model, our method achievesstate-of-the-art performance on the BOP benchmark, surpassing previousmethods with minimal fine-tuning.

ACCV

ICANN

RDGE-6D: Reverse Direction Geometry Injection for 6D Pose Estimation
- 分类：实例级方法
- 论文：https://link.springer.com/chapter/10.1007/978-3-032-04546-1_25
- 代码：代码未开源
- 摘要：The use of RGBD images for obtaining high-accuracy 6D poseestimation results has gained popularity due to the availability ofinexpensive RGB-D sensors. Typically, state-of-the-art methods employdeep learning-based methods to process RGBD data for pose estimation.They either use different backbones to extract features from each typeof image or combine preprocessed depth images with RGB images beforefeeding them into a network. However, with increased network layers,deep networks tend to lose detail, which may be crucial for objectlocalization. Additionally, the influence of occlusion also becomes morepronounced, posing challenges in effectively focusing on target objects.To address these issues, we propose RDGE-6D, an intuitive and effectivemethod that leverages reverse direction fusion to enhance RGB deepfeatures using shallow features from depth images. Extensive experimentson benchmark datasets demonstrate the superior performance of ourmethod, especially in scenarios with heavy occlusion.

ACML

其他

WACV

Generalizable Single-View Object Pose Estimation by Two-SideGenerating and Matching
- 分类：任意物体方法
- 论文：https://openaccess.thecvf.com/content/WACV2025/html/Sun_Generalizable_Single-View_Object_Pose_Estimation_by_Two-Side_Generating_and_Matching_WACV_2025_paper.html
- 代码：https://github.com/scy639/Gen2SM
- 摘要：In this paper, we present a novel generalizable object poseestimation method to determine the object pose using only one RGB image.Unlike traditional approaches that rely on instance-level object poseestimation and necessitate extensive training data, our method offersgeneralization to unseen objects without extensive training, operateswith a single reference image of the object, and eliminates the need for3D object models or multiple views of the object. These characteristicsare achieved by utilizing a diffusion model to generate novel-viewimages and conducting a two-sided matching on these generated images.Quantitative experiments demonstrate the superiority of our method overexisting pose estimation techniques across both synthetic and real-worlddatasets. Remarkably, our approach maintains strong performance even inscenarios with significant viewpoint changes, highlighting itsrobustness and versatility in challenging conditions. The code will bereleased at https://github.com/scy639/Gen2SM.
Shape-Biased Texture Agnostic Representations for ImprovedTextureless and Metallic Object Detection and 6D Pose Estimation
- 分类：实例级方法
- 论文：https://openaccess.thecvf.com/content/WACV2025/html/Honig_Shape-Biased_Texture_Agnostic_Representations_for_Improved_Textureless_and_Metallic_Object_WACV_2025_paper.html
- 代码：https://github.com/hoenigpeter/randomized_texturing
- 摘要：Recent advances in machine learning have greatly benefitedobject detection and 6D pose estimation. However, textureless andmetallic objects still pose a significant challenge due to few visualcues and the texture bias of CNNs. To address this issue, we propose astrategy for inducing a shape bias to CNN training. In particular, byrandomizing textures applied to object surfaces during data rendering,we create training data without consistent texture cues. Thismethodology allows for seamless integration into existing data renderingengines, and results in negligible computational overhead for datarendering and network training. Our findings demonstrate that the shapebias we induce via randomized texturing, improves over existingapproaches using style transfer. We evaluate with five detectors and twopose estimators. For three object detectors and for pose estimation ingeneral, estimation accuracy improves for textureless and metallicobjects. Additionally we show that our approach increases the poseestimation accuracy in the presence of image noise and strongillumination changes. Code available at https://github.com/hoenigpeter/randomized_texturing.

CoRL

One View, Many Worlds: Single-Image to 3D object Meets GenerativeDomain Randomization for One-Shot 6D Pose Estimation
- 分类：任意物体方法
- 论文：https://openreview.net/pdf?id=kto4zVmo4w
- 代码：https://github.com/gzwsama/OnePoseviaGen
- 摘要：Estimating the 6D pose of arbitrary objects from a singlereference image is a critical yet challenging task in robotics,especially considering the long-tail distribution of real-worldinstances. While category-level and model-based approaches have achievednotable progress, they remain limited in generalizing to unseen objectsunder one-shot settings. In this work, we propose a novel pipeline forfast and accurate one-shot 6D pose and scale estimation. Leveragingrecent advances in single-view 3D generation, we first buildhigh-fidelity textured meshes without requiring known object poses. Toresolve scale ambiguity, we introduce a coarse-to-fine alignment modulethat estimates both object size and initial pose by matching 2D-3Dfeatures with depth information. We then generate a diversified set ofplausible 3D models using text-guided generative augmentation and renderthem with Blender to synthesize large-scale, domain-randomized trainingdata for pose estiamtion. This synthetic data bridges the domain gap andenables robust fine-tuning of pose estimators. Our method achievesstate-of-the-art results on several 6D pose benchmarks, and we furthervalidate its effectiveness on a newly collected in-the-wild dataset.Finally, we integrate our system with a dexterous hand, demonstratingits robustness in real-world robotic grasping tasks. All code, data, andmodels will be released to foster future research.
PicoPose: Progressive Pixel-to-Pixel Correspondence Learning forNovel Object Pose Estimation
- 分类：任意物体方法
- 论文：https://openreview.net/pdf?id=EXgckdYESp
- 代码：https://github.com/foollh/PicoPose
- 摘要：RGB-based novel object pose estimation is critical for rapiddeployment in robotic applications, yet zero-shot generalization remainsa key challenge. In this paper, we introduce PicoPose, a novel frameworkdesigned to tackle this task using a three-stage pixel-to-pixelcorrespondence learning process. Firstly, PicoPose matches features fromthe RGB observation with those from rendered object templates,identifying the best-matched template and establishing coarsecorrespondences. Secondly, PicoPose smooths the correspondences byglobally regressing a 2D affine transformation, including in-planerotation, scale, and 2D translation, from the coarse correspondence map.Thirdly, PicoPose applies the affine transformation to the feature mapof the best-matched template and learns correspondence offsets withinlocal regions to achieve fine-grained correspondences. By progressivelyrefining the correspondences, PicoPose significantly improves theaccuracy of object poses computed via PnP/RANSAC. PicoPose achievesstate-of-the-art performance on the seven core datasets of the BOPbenchmark, demonstrating exceptional generalization to novel objects.Code and trained models are available at https://github.com/foollh/PicoPose.
UnPose: Uncertainty-Guided Diffusion Priors for Zero-Shot PoseEstimation
- 分类：任意物体方法
- 论文：https://openreview.net/pdf?id=L5PJBd8ahD
- 代码：代码未开源
- 摘要：Estimating the 6D pose of novel objects is a fundamental yetchallenging problem in robotics, often relying on access to object CADmodels. However, acquiring such models can be costly and impractical.Recent approaches aim to bypass this requirement by leveraging strongpriors from foundation models to reconstruct objects from single ormulti-view images, but typically require additional training or producehallucinated geometry. To this end, we propose UnPose, a novel frameworkfor zero-shot, model-free 6D object pose estimation and reconstructionthat exploits 3D priors and uncertainty estimates from a pre-traineddiffusion model. Specifically, starting from a single-view RGB-D frame,UnPose uses a multi-view diffusion model to estimate an initial 3D modelusing 3D Gaussian Splatting (3DGS) representation, along with pixel-wiseepistemic uncertainty estimates. As additional observations becomeavailable, we incrementally refine the 3DGS model by fusing new viewsguided by the diffusion model’s uncertainty, thereby, continuouslyimproving the pose estimation accuracy and 3D reconstruction quality. Toensure global consistency, the diffusion prior-generated views andsubsequent observations are further integrated in a pose graph andjointly optimized into a coherent 3DGS field. Extensive experimentsdemonstrate that UnPose significantly outperforms existing approaches inboth 6D pose estimation accuracy and 3D reconstruction quality. Wefurther showcase its practical applicability in real-world roboticmanipulation tasks. Video demos can be found at our project page: https://frankzhaodong.github.io/UnPose.

RSS

V-HOP: Visuo-Haptic 6D Object Pose Tracking
- 分类：任意物体跟踪方法
- 论文：https://roboticsconference.org/2025/program/papers/37/
- 代码：https://github.com/brown-ivl/v-hop
- 摘要：Humans naturally integrate vision and haptics for robustobject perception during manipulation. The loss of either modalitysignificantly degrades performance. Inspired by this multisensoryintegration, prior object pose estimation research has attempted tocombine visual and haptic/tactile feedback. Although these worksdemonstrate improvements in controlled environments or syntheticdatasets, they often underperform vision-only approaches in real-worldsettings due to poor generalization across diverse grippers, sensorlayouts, or sim-to-real environments. Furthermore, they typicallyestimate the object pose for each frame independently, resulting in lesscoherent tracking over sequences in real-world deployments. To addressthese limitations, we introduce a novel unified haptic representationthat effectively handles multiple gripper embodiments. Building on thisrepresentation, we introduce a new visuo-haptic transformer-based objectpose tracker that seamlessly integrates visual and haptic input. Wevalidate our framework in our dataset and the Feelsight dataset,demonstrating significant performance improvement on challengingsequences. Notably, our method achieves superior generalization androbustness across novel embodiments, objects, and sensor types (bothtactile-based and vision-based tactile sensors). In real-worldexperiments, we demonstrate that our approach outperformsstate-of-the-art visual trackers by a large margin. We further show thatwe can achieve precise manipulation tasks by incorporating our real-timeobject tracking result into motion plans, underscoring the advantages ofvisuo-haptic perception. Project website: https://ivl.cs.brown.edu/research/v-hop.

SIGGRAPH Asia

【环境配置】tmux配置文件

2026-06-12T15:29:13.000Z

在家目录下，新建一个文件名为.tmux.conf的文件，将下面的内容写入文件，即可完成配置。

set-option -g status-keys vi
setw -g mode-keys vi

setw -g monitor-activity on

# setw -g c0-change-trigger 10
# setw -g c0-change-interval 100

# setw -g c0-change-interval 50
# setw -g c0-change-trigger  75

set-window-option -g automatic-rename on
set-option -g set-titles on
set -g history-limit 100000

# set-window-option -g utf8 on

# set command prefix
set-option -g prefix C-a
unbind-key C-b
bind-key C-a send-prefix

bind h select-pane -L
bind j select-pane -D
bind k select-pane -U
bind l select-pane -R

bind -n M-Left select-pane -L
bind -n M-Right select-pane -R
bind -n M-Up select-pane -U
bind -n M-Down select-pane -D

bind < resize-pane -L 7
bind > resize-pane -R 7
bind - resize-pane -D 7
bind + resize-pane -U 7

bind-key -n M-l next-window
bind-key -n M-h previous-window

set -g status-interval 1
# status bar
set -g status-bg black
set -g status-fg blue

# set -g status-utf8 on
set -g status-justify centre
set -g status-bg default
set -g status-left " #[fg=green]#S@#H #[default]"
set -g status-left-length 20

# mouse support
# for tmux 2.1
# set -g mouse-utf8 on
set -g mouse on

# for previous version
# set -g mode-mouse on
# set -g mouse-resize-pane on
# set -g mouse-select-pane on
# set -g mouse-select-window on

# set -g status-right-length 25
set -g status-right "#[fg=green]%H:%M:%S #[fg=magenta]%a %m-%d #[default]"

# fix for tmux 1.9
bind '"' split-window -vc "#{pane_current_path}"
bind '%' split-window -hc "#{pane_current_path}"
bind 'c' new-window -c "#{pane_current_path}"

# run-shell "powerline-daemon -q"

# vim: ft=conf

【环境配置】vim配置文件

2026-06-12T15:28:42.000Z

在家目录下，新建一个文件名为.vimrc的文件，将下面的内容写入文件，即可完成配置。

" An example for a vimrc file.
"
" To use it, copy it to
"   for Unix and OS/2: ~/.vimrc
"   for Amiga: s:.vimrc
"   for MS-DOS and Win32: $VIM\_vimrc
"   for OpenVMS: sys$login:.vimrc

" When started as "evim", evim.vim will already have done these settings.
if v:progname =~? "evim"
  finish
endif

" Use Vim settings, rather then Vi settings (much better!).
" This must be first, because it changes other options as a side effect.
set nocompatible

" allow backspacing over everything in insert mode
set backspace=indent,eol,start

if has("vms")
  set nobackup  " do not keep a backup file, use versions instead
else
  set backup    " keep a backup file
endif
set history=50  " keep 50 lines of command line history
set ruler       " show the cursor position all the time
set showcmd     " display incomplete commands
set incsearch   " do incremental searching
"==========================================================================
" My Setting
"==========================================================================
vmap y :w! /tmp/vitmp
nmap p :r! cat /tmp/vitmp

" 语法高亮
syntax enable
syntax on
" 显示行号
set nu

" 修改默认注释颜色
" hi Comment ctermfg=DarkCyan
" 允许退格键删除
" set backspace=2
" 启用鼠标
set mouse=a
set selection=exclusive
set selectmode=mouse,key
" 按C语言格式缩进
set cindent
set autoindent
set smartindent
set shiftwidth=4

" 允许在有未保存的修改时切换缓冲区
" set hidden

" 设置无备份文件
set writebackup
set nobackup

" 显示括号匹配
set showmatch
" 括号匹配显示时间为1(单位是十分之一秒)
set matchtime=5
" 显示当前的行号列号：
set ruler
" 在状态栏显示正在输入的命令
set showcmd

set foldmethod=syntax
" 默认情况下不折叠
set foldlevel=100
" 开启状态栏信息
set laststatus=2
" 命令行的高度，默认为1，这里设为2
set cmdheight=2


" 显示Tab符，使用一高亮竖线代替
set list
" set listchars=tab:\|\ ,
set listchars=tab:>-,trail:-


" 侦测文件类型
filetype on
" 载入文件类型插件
filetype plugin on
" 为特定文件类型载入相关缩进文件
filetype indent on
" 启用自动补全
filetype plugin indent on


" 设置编码自动识别，中文引号显示
filetype on " 打开文件类型检测
" set fileencodings=euc-cn,ucs-bom,utf-8,cp936,gb2312,gb18030,gbk,big5,euc-jp,euc-kr,latin1
set fileencodings=utf-8,gb2312,gbk,gb18030
" 这个功能很给劲，不管encoding是什么编码，都能将文本显示汉字
" set termencoding=gb2312
set termencoding=utf-8
" 新建文件使用的编码
set fileencoding=utf-8
" set fileencoding=gb2312
" 用于显示的编码，仅仅是显示
set encoding=utf-8
" set encoding=utf-8
" set encoding=euc-cn
" set encoding=gbk
" set encoding=gb2312
" set ambiwidth=double
set fileformat=unix


" 设置高亮搜索
set hlsearch
" 在搜索时，输入的词句的逐字符高亮
set incsearch

" 着色模式
set t_Co=256
" colorscheme wombat256mod
" colorscheme gardener
" colorscheme elflord
colorscheme desert
" colorscheme evening
" colorscheme darkblue
" colorscheme torte
" colorscheme default

" 字体 && 字号
" set guifont=Monaco:h10
set guifont=Consolas:h10

" ======= 引号 && 括号自动匹配 ======= "
"
" :inoremap ( ()i

" :inoremap ) =ClosePair(')')
"
" :inoremap { {}i
"
" :inoremap } =ClosePair('}')
"
" :inoremap [ []i
"
" :inoremap ] =ClosePair(']')
"
" :inoremap < <>i
"
" :inoremap > =ClosePair('>')
"
" :inoremap " ""i
"
" :inoremap ' ''i
"
" :inoremap ` ``i
"
" :inoremap * **i

" 每行超过80个的字符用下划线标示
" au BufRead,BufNewFile *.s,*.asm,*.h,*.c,*.cpp,*.java,*.cs,*.lisp,*.el,*.erl,*.tex,*.sh,*.lua,*.pl,*.php,*.tpl,*.py,*.rb,*.erb,*.vim,*.js,*.jade,*.coffee,*.css,*.xml,*.html,*.shtml,*.xhtml Underlined /.\%81v/
"
"
" For Win32 GUI: remove 't' flag from 'guioptions': no tearoff menu entries
" let &guioptions = substitute(&guioptions, "t", "", "g")

" Don't use Ex mode, use Q for formatting
map Q gq

" This is an alternative that also works in block mode, but the deleted
" text is lost and it only works for putting the current register.
" vnoremap p "_dp

" Switch syntax highlighting on, when the terminal has colors
" Also switch on highlighting the last used search pattern.
if &t_Co > 2 || has("gui_running")
  syntax on
  set hlsearch
endif

" Only do this part when compiled with support for autocommands.
if has("autocmd")

  " Enable file type detection.
  " Use the default filetype settings, so that mail gets 'tw' set to 72,
  " 'cindent' is on in C files, etc.
  " Also load indent files, to automatically do language-dependent indenting.
  filetype plugin indent on

  " Put these in an autocmd group, so that we can delete them easily.
  augroup vimrcEx
  au!

  " For all text files set 'textwidth' to 80 characters.
  autocmd FileType text setlocal textwidth=80

  " When editing a file, always jump to the last known cursor position.
  " Don't do it when the position is invalid or when inside an event handler
  " (happens when dropping a file on gvim).
  autocmd BufReadPost *
    \ if line("'\"") > 0 && line("'\"") <= line("$") |
    \   exe "normal g`\"" |
    \ endif

  augroup END

else

  set autoindent " always set autoindenting on

endif " has("autocmd")

" 增加鼠标行高亮
set cursorline
hi CursorLine cterm=NONE ctermbg=darkred ctermfg=white

" 设置tab是四个空格
set ts=4
set expandtab

" 主要给Tlist使用
let Tlist_Exit_OnlyWindow = 1
let Tlist_Auto_Open = 1

Hello World

2026-06-11T16:00:00.000Z

Welcome to Hexo! This is your veryfirst post. Check documentation formore info. If you get any problems when using Hexo, you can find theanswer in troubleshooting oryou can ask me on GitHub.

Quick Start

Create a new post

1	$ hexo new "My New Post"

More info: Writing

Run server

1	$ hexo server

More info: Server

Generate static files

1	$ hexo generate

More info: Generating

Deploy to remote sites

1	$ hexo deploy

More info: Deployment