【代码复现】Instance-Adaptive and Geometric-Aware Keypoint Learning for Category-Level 6D Object Pose Estimation

博主： karltan0328
发布时间：2025 年 12 月 16 日
10 次浏览
暂无评论
11837字数
分类：代码复现

### 安装AG-Pose环境

1. 创建环境：
   
   ```bash
   conda create --name agpose python=3.9
   ```
2. 激活环境：
   
   ```bash
   conda activate agpose
   ```
3. 安装PyTorch：
   
   ```bash
   pip3 install torch==1.12 torchvision torchaudio --index-url https://download.pytorch.org/whl/cu113
   ```
4. 安装其他依赖：
   
   ```bash
   pip3 install gorilla-core==0.2.5.3
   pip3 install opencv-python
   
   git clone https://github.com/Leeiieeo/AG-Pose.git
   cd AG-Pose/model/pointnet2
   python setup.py install
   ```
5. Clone [mentian/object-deformnet](https://github.com/mentian/object-deformnet)并安装`nn_distance`，需要使用该仓库中的代码进行数据处理：
   
   ```bash
   git clone https://github.com/mentian/object-deformnet.git
   cd object-deformnet/lib/nn_distance
   python setup.py install
   ```

### 下载数据集

1. 下载NOCS数据集到数据集文件夹中：
   
   - [http://download.cs.stanford.edu/orion/nocs/camera_composed_depth.zip](http://download.cs.stanford.edu/orion/nocs/camera_composed_depth.zip)
   - [http://download.cs.stanford.edu/orion/nocs/camera_train.zip](http://download.cs.stanford.edu/orion/nocs/camera_train.zip)
   - [http://download.cs.stanford.edu/orion/nocs/camera_val25K.zip](http://download.cs.stanford.edu/orion/nocs/camera_val25K.zip)
   - [http://download.cs.stanford.edu/orion/nocs/gts.zip](http://download.cs.stanford.edu/orion/nocs/gts.zip)
   - [http://download.cs.stanford.edu/orion/nocs/obj_models.zip](http://download.cs.stanford.edu/orion/nocs/obj_models.zip)
   - [http://download.cs.stanford.edu/orion/nocs/real_test.zip](http://download.cs.stanford.edu/orion/nocs/real_test.zip)
   - [http://download.cs.stanford.edu/orion/nocs/real_train.zip](http://download.cs.stanford.edu/orion/nocs/real_train.zip)
2. 下载[JiehongLin/Self-DPDN](https://github.com/JiehongLin/Self-DPDN)的分割结果到仓库文件夹中：
   
   - [https://drive.google.com/file/d/1hNmNRr7YRCgg-c_qdvaIzKEd2g4Kac3w/](https://drive.google.com/file/d/1hNmNRr7YRCgg-c_qdvaIzKEd2g4Kac3w/)
3. 下载使用[mentian/object-deformnet](https://github.com/mentian/object-deformnet)进行数据预处理时用到的数据到仓库文件夹中：
   
   - [https://drive.google.com/file/d/1p72NdY4Bie_sra9U8zoUNI4fTrQZdbnc/](https://drive.google.com/file/d/1p72NdY4Bie_sra9U8zoUNI4fTrQZdbnc/)
4. 下载运行AG-Pose时使用到的分割结果：
   
   - [http://home.ustc.edu.cn/~llinxiao/segmentation_results.zip](http://home.ustc.edu.cn/~llinxiao/segmentation_results.zip)
5. 下载NOCS数据集中缺失的物体模型（From: [https://github.com/mentian/object-deformnet/issues/3#issuecomment-698858332](https://github.com/mentian/object-deformnet/issues/3#issuecomment-698858332)）到数据集文件夹中：
   
   - [https://drive.google.com/file/d/1rWkxEVJJh_kWIqxudn_i6sJ-hhd0E7TV/](https://drive.google.com/file/d/1rWkxEVJJh_kWIqxudn_i6sJ-hhd0E7TV/)

### 处理数据集

#### JiehongLin/Self-DPDN

遵循[JiehongLin/Self-DPDN](https://github.com/JiehongLin/Self-DPDN)的处理方式，首先解压上述文件，并组织如下，注意，data文件夹位于Self-DPDN文件夹下：

```text
data/
├── camera -> /data1/dataset/nocs/tanmx/dpdn/camera
│   ├── train
│   └── val
├── camera_full_depths -> /data1/dataset/nocs/tanmx/dpdn/camera_full_depths/
│   ├── train
│   └── val
├── gts -> /data1/dataset/nocs/tanmx/dpdn/gts/
│   ├── real_test
│   └── val
├── mean_shapes.npy
├── obj_models -> /data1/dataset/nocs/tanmx/dpdn/obj_models/
│   ├── real_test
│   ├── real_train
│   ├── train
│   └── val
├── real -> /data1/dataset/nocs/tanmx/dpdn/real/
│   ├── test
│   └── train
└── segmentation_results -> /data1/dataset/nocs/tanmx/dpdn/segmentation_results/
    ├── test_trainedwithMask
    ├── test_trainedwoMask
    └── train_trainedwoMask

21 directories, 1 file
```

进入`obj_models/val/02876657/d3b53f56b4a7b3b3c9f016d57db96408`文件夹，其中内容为

```bash
$ ls
bbox.txt  model.mtl  model.obj
```

查看`model.obj`文件：

```bash
$ cat model.obj
newmtl Material.57774a50243e5d429f3a4ea105ec3930
d 0.5
Tr 0.5
Kd 0.392157 0.584314 0.929412
newmtl Material.90b00bf69165eda831fe702671104217
Tr 0.0
Kd 0.6 0.117647 0.117647
newmtl Material.dc7a3ca9ca091ddbf48733c3a604f557
Tr 0.0
Kd 1.0 1.0 1.0
newmtl Material.b2cdaf4394d1a4f850458878b81c7fa1
Tr 0.0
Kd 0.8 0.6 0.0
```

解压[https://drive.google.com/file/d/1rWkxEVJJh_kWIqxudn_i6sJ-hhd0E7TV/](https://drive.google.com/file/d/1rWkxEVJJh_kWIqxudn_i6sJ-hhd0E7TV/)中的文件：

```bash
$ unzip obj_models.zip
$ ls
02876657  obj_models.zip
$ ls 02876657/d3b53f56b4a7b3b3c9f016d57db96408/
bbox.txt  info.txt  model_bad.mtl  model_bad.obj  model.mtl  model.obj
```

删除原文件夹`obj_models/val/02876657/d3b53f56b4a7b3b3c9f016d57db96408`中的内容，将`02876657/d3b53f56b4a7b3b3c9f016d57db96408`文件夹中的内容复制到该文件夹中：

```bash
$ rm obj_models/val/02876657/d3b53f56b4a7b3b3c9f016d57db96408/*
$ cp 02876657/d3b53f56b4a7b3b3c9f016d57db96408/* obj_models/val/02876657/d3b53f56b4a7b3b3c9f016d57db96408/
$ ls obj_models/val/02876657/d3b53f56b4a7b3b3c9f016d57db96408
bbox.txt  info.txt  model_bad.mtl  model_bad.obj  model.mtl  model.obj
```

执行：

```bash
python data_processing.py
```

执行后目录变为：

```text
data/
├── camera -> /data1/dataset/nocs/tanmx/dpdn/camera
│   ├── train
│   ├── train_list_all.txt
│   ├── train_list.txt
│   ├── val
│   └── val_list_all.txt
├── camera_full_depths -> /data1/dataset/nocs/tanmx/dpdn/camera_full_depths/
│   ├── train
│   └── val
├── gts -> /data1/dataset/nocs/tanmx/dpdn/gts/
│   ├── real_test
│   └── val
├── mean_shapes.npy
├── obj_models -> /data1/dataset/nocs/tanmx/dpdn/obj_models/
│   ├── real_test
│   ├── real_train
│   ├── train
│   └── val
├── real -> /data1/dataset/nocs/tanmx/dpdn/real/
│   ├── test
│   ├── test_list_all.txt
│   ├── train
│   ├── train_list_all.txt
│   └── train_list.txt
└── segmentation_results -> /data1/dataset/nocs/tanmx/dpdn/segmentation_results
    ├── test_trainedwithMask
    ├── test_trainedwoMask
    └── train_trainedwoMask

21 directories, 7 files
```

执行：

```bash
$ ls -l data/camera/
total 9712
drwxr-xr-x 27502 tanmx tanmx  577536 Jun 14  2019 train
-rw-rw-r--     1 tanmx tanmx 4675000 Apr 29 11:00 train_list_all.txt
-rw-rw-r--     1 tanmx tanmx 4235159 Apr 29 15:57 train_list.txt
drwxr-xr-x  2502 tanmx tanmx   69632 Jun 14  2019 val
-rw-rw-r--     1 tanmx tanmx  375000 Apr 29 11:00 val_list_all.txt

$ ls -l data/real/
total 228
drwxrwxr-x 8 tanmx tanmx  4096 Nov 13  2018 test
-rw-rw-r-- 1 tanmx tanmx 49572 Apr 29 11:00 test_list_all.txt
drwxrwxr-x 9 tanmx tanmx  4096 Jun 14  2019 train
-rw-rw-r-- 1 tanmx tanmx 82042 Apr 29 11:00 train_list_all.txt
-rw-rw-r-- 1 tanmx tanmx 82042 Apr 29 16:09 train_list.txt
```

#### mentian/object-deformnet

遵循[mentian/object-deformnet](https://github.com/mentian/object-deformnet)的处理方式，首先解压上述文件，并组织如下，注意，data文件夹位于object-deformnet文件夹下：

```text
data/
├── camera -> /data1/dataset/nocs/tanmx/spd/camera/
│   ├── train
│   └── val
├── deformnet_eval -> /data1/dataset/nocs/tanmx/spd/deformnet_eval/
│   ├── camera
│   ├── mrcnn_results
│   ├── nocs_results
│   └── real
├── gts -> /data1/dataset/nocs/tanmx/spd/gts/
│   ├── real_test
│   └── val
├── obj_models -> /data1/dataset/nocs/tanmx/spd/obj_models/
│   ├── real_test
│   ├── real_train
│   ├── train
│   └── val
├── pose_dataset.py
├── real -> /data1/dataset/nocs/tanmx/spd/real/
│   ├── test
│   └── train
└── shape_dataset.py

19 directories, 2 files
```

注意，这里也要和替换掉缺失的物体模型。

**按顺序**执行：

```bash
cd preprocess
python shape_data.py
python pose_data.py
```

执行`python shape_data.py`后目录变为：

```text
data/
├── camera -> /data1/dataset/nocs/tanmx/spd/camera/
│   ├── train
│   └── val
├── deformnet_eval -> /data1/dataset/nocs/tanmx/spd/deformnet_eval/
│   ├── camera
│   ├── mrcnn_results
│   ├── nocs_results
│   └── real
├── gts -> /data1/dataset/nocs/tanmx/spd/gts/
│   ├── real_test
│   └── val
├── obj_models -> /data1/dataset/nocs/tanmx/spd/obj_models/
│   ├── camera_train.pkl
│   ├── camera_val.pkl
│   ├── mug_meta.pkl
│   ├── real_test
│   ├── real_test.pkl
│   ├── real_train
│   ├── real_train.pkl
│   ├── ShapeNetCore_2048.h5
│   ├── ShapeNetCore_4096.h5
│   ├── train
│   └── val
├── pose_dataset.py
├── real -> /data1/dataset/nocs/tanmx/spd/real/
│   ├── test
│   └── train
└── shape_dataset.py

19 directories, 9 files
```

执行`python pose_data.py`后目录变为：

```text
data/
├── camera -> /data1/dataset/nocs/tanmx/spd/camera
│   ├── train
│   ├── train_list_all.txt
│   ├── train_list.txt
│   ├── val
│   ├── val_list_all.txt
│   └── val_list.txt
├── deformnet_eval -> /data1/dataset/nocs/tanmx/spd/deformnet_eval
│   ├── camera
│   ├── mrcnn_results
│   ├── nocs_results
│   └── real
├── gts -> /data1/dataset/nocs/tanmx/spd/gts/
│   ├── real_test
│   └── val
├── obj_models -> /data1/dataset/nocs/tanmx/spd/obj_models/
│   ├── camera_train.pkl
│   ├── camera_val.pkl
│   ├── mug_meta.pkl
│   ├── real_test
│   ├── real_test.pkl
│   ├── real_train
│   ├── real_train.pkl
│   ├── ShapeNetCore_2048.h5
│   ├── ShapeNetCore_4096.h5
│   ├── train
│   └── val
├── pose_dataset.py
├── real -> /data1/dataset/nocs/tanmx/spd/real/
│   ├── test
│   ├── test_list_all.txt
│   ├── test_list.txt
│   ├── train
│   ├── train_list_all.txt
│   └── train_list.txt
└── shape_dataset.py

19 directories, 17 files
```

执行：

```bash
$ ls -l data/camera/
total 9996
drwxr-xr-x 27502 tanmx tanmx  577536 Jun 14  2019 train
-rw-rw-r--     1 tanmx tanmx 4675000 Apr 29 14:00 train_list_all.txt
-rw-rw-r--     1 tanmx tanmx 4235159 Apr 29 18:42 train_list.txt
drwxr-xr-x  2502 tanmx tanmx   69632 Jun 14  2019 val
-rw-rw-r--     1 tanmx tanmx  375000 Apr 29 14:00 val_list_all.txt
-rw-rw-r--     1 tanmx tanmx  288300 Apr 29 19:15 val_list.txt

$ ls -l data/real/
total 280
drwxrwxr-x 8 tanmx tanmx  4096 Nov 13  2018 test
-rw-rw-r-- 1 tanmx tanmx 49572 Apr 29 14:00 test_list_all.txt
-rw-rw-r-- 1 tanmx tanmx 49572 Apr 29 19:21 test_list.txt
drwxrwxr-x 9 tanmx tanmx  4096 Jun 14  2019 train
-rw-rw-r-- 1 tanmx tanmx 82042 Apr 29 14:00 train_list_all.txt
-rw-rw-r-- 1 tanmx tanmx 82042 Apr 29 18:53 train_list.txt

$ ls -l data/obj_models/
total 116060
-rw-rw-r--  1 tanmx tanmx 26695895 Apr 29 11:32 camera_train.pkl
-rw-rw-r--  1 tanmx tanmx  4634293 Apr 29 11:36 camera_val.pkl
-rw-rw-r--  1 tanmx tanmx    23197 Apr 29 11:37 mug_meta.pkl
drwxrwxr-x  2 tanmx tanmx     4096 Sep 27  2019 real_test
-rw-rw-r--  1 tanmx tanmx   443600 Apr 29 11:37 real_test.pkl
drwxrwxr-x  2 tanmx tanmx     4096 Sep 27  2019 real_train
-rw-rw-r--  1 tanmx tanmx   443584 Apr 29 11:37 real_train.pkl
-rw-rw-r--  1 tanmx tanmx 28950893 Apr 29 12:21 ShapeNetCore_2048.h5
-rw-rw-r--  1 tanmx tanmx 57618690 Apr 29 11:39 ShapeNetCore_4096.h5
drwxr-xr-x 12 tanmx tanmx     4096 Sep 27  2019 train
drwxr-xr-x 12 tanmx tanmx     4096 Sep 27  2019 val
```

### 运行代码

需要[JiehongLin/Self-DPDN](https://github.com/JiehongLin/Self-DPDN)处理后的数据集及其list和[mentian/object-deformnet](https://github.com/mentian/object-deformnet)处理后的物体模型pkl，组织如下：

```text
data/
├── camera -> /data1/dataset/nocs/tanmx/agpose/camera
│   ├── train
│   ├── train_list_all.txt
│   ├── train_list.txt
│   ├── val
│   └── val_list_all.txt
├── camera_full_depths -> /data1/dataset/nocs/tanmx/agpose/camera_full_depths/
│   ├── train
│   └── val
├── gts -> /data1/dataset/nocs/tanmx/agpose/gts/
│   ├── real_test
│   └── val
├── obj_models -> /data1/dataset/nocs/tanmx/agpose/obj_models/
│   ├── camera_train.pkl
│   ├── real_test
│   ├── real_train
│   ├── real_train.pkl
│   ├── train
│   └── val
├── real -> /data1/dataset/nocs/tanmx/agpose/real/
│   ├── test
│   ├── test_list_all.txt
│   ├── train
│   ├── train_list_all.txt
│   └── train_list.txt
└── segmentation_results -> /data1/dataset/nocs/tanmx/agpose/segmentation_results
    ├── CAMERA25
    └── REAL275

20 directories, 8 files
```

之后按照仓库README中的命令运行代码即可。

最后修改：2025 年 12 月 16 日

如果觉得我的文章对你有用，请随意赞赏。

发表评论取消回复
使用cookie技术保留您的个人信息以便您下次快速评论，继续评论表示您已同意该条款。

评论 *

私密评论

名称 *

🎲

邮箱 *

地址

【代码复现】Instance-Adaptive and Geometric-Aware Keypoint Learning for Category-Level 6D Object Pose Estimation

karltan0328 • 2025 年 12 月 16 日

### 安装AG-Pose环境

### 下载数据集

### 处理数据集

#### JiehongLin/Self-DPDN

遵循[JiehongLin/Self-DPDN](https://github.com/JiehongLin/Self-DPDN)的处理方式，首先解压上述文件，并组织如下，注意，data文件夹位于Self-DPDN文件夹下：

21 directories, 1 file
```

进入`obj_models/val/02876657/d3b53f56b4a7b3b3c9f016d57db96408`文件夹，其中内容为

```bash
$ ls
bbox.txt  model.mtl  model.obj
```

查看`model.obj`文件：

解压[https://drive.google.com/file/d/1rWkxEVJJh_kWIqxudn_i6sJ-hhd0E7TV/](https://drive.google.com/file/d/1rWkxEVJJh_kWIqxudn_i6sJ-hhd0E7TV/)中的文件：

```bash
$ unzip obj_models.zip
$ ls
02876657  obj_models.zip
$ ls 02876657/d3b53f56b4a7b3b3c9f016d57db96408/
bbox.txt  info.txt  model_bad.mtl  model_bad.obj  model.mtl  model.obj
```

删除原文件夹`obj_models/val/02876657/d3b53f56b4a7b3b3c9f016d57db96408`中的内容，将`02876657/d3b53f56b4a7b3b3c9f016d57db96408`文件夹中的内容复制到该文件夹中：

执行：

```bash
python data_processing.py
```

执行后目录变为：

21 directories, 7 files
```

执行：

#### mentian/object-deformnet

遵循[mentian/object-deformnet](https://github.com/mentian/object-deformnet)的处理方式，首先解压上述文件，并组织如下，注意，data文件夹位于object-deformnet文件夹下：

19 directories, 2 files
```

注意，这里也要和替换掉缺失的物体模型。

**按顺序**执行：

```bash
cd preprocess
python shape_data.py
python pose_data.py
```

执行`python shape_data.py`后目录变为：

```text
data/
├── camera -> /data1/dataset/nocs/tanmx/spd/camera/
│   ├── train
│   └── val
├── deformnet_eval -> /data1/dataset/nocs/tanmx/spd/deformnet_eval/
│   ├── camera
│   ├── mrcnn_results
│   ├── nocs_results
│   └── real
├── gts -> /data1/dataset/nocs/tanmx/spd/gts/
│   ├── real_test
│   └── val
├── obj_models -> /data1/dataset/nocs/tanmx/spd/obj_models/
│   ├── camera_train.pkl
│   ├── camera_val.pkl
│   ├── mug_meta.pkl
│   ├── real_test
│   ├── real_test.pkl
│   ├── real_train
│   ├── real_train.pkl
│   ├── ShapeNetCore_2048.h5
│   ├── ShapeNetCore_4096.h5
│   ├── train
│   └── val
├── pose_dataset.py
├── real -> /data1/dataset/nocs/tanmx/spd/real/
│   ├── test
│   └── train
└── shape_dataset.py

19 directories, 9 files
```

执行`python pose_data.py`后目录变为：

19 directories, 17 files
```

执行：

### 运行代码

20 directories, 8 files
```

之后按照仓库README中的命令运行代码即可。

【代码复现】Instance-Adaptive and Geometric-Aware Keypoint Learning for Category-Level 6D Object Pose Estimation

发表评论取消回复
使用cookie技术保留您的个人信息以便您下次快速评论，继续评论表示您已同意该条款。

欢迎使用Typecho

【环境配置】vim配置文件

【环境配置】tmux配置文件

【论文列表】2025年Object Pose Estimation顶会论文列表及分类

【代码复现】Instance-Adaptive and Geometric-Aware Keypoint Learning for Category-Level 6D Object Pose Estimation

【环境配置】tmux配置文件

【代码复现】OnePose: One-Shot Object Pose Estimation without CAD Models

欢迎使用Typecho

【环境配置】vim配置文件

【代码复现】Instance-Adaptive and Geometric-Aware Keypoint Learning for Category-Level 6D Object Pose Estimation

【代码复现】Instance-Adaptive and Geometric-Aware Keypoint Learning for Category-Level 6D Object Pose Estimation

发表评论 取消回复 使用cookie技术保留您的个人信息以便您下次快速评论，继续评论表示您已同意该条款。

【代码复现】Instance-Adaptive and Geometric-Aware Keypoint Learning for Category-Level 6D Object Pose Estimation

发表评论取消回复
使用cookie技术保留您的个人信息以便您下次快速评论，继续评论表示您已同意该条款。