Omni6D: Large-Vocabulary 3D Object Dataset for Category-Level 6D Object Pose Estimation

Mengchen Zhang  Tong Wu  Tai Wang  Tengfei Wang  Ziwei Liu  Dahua Lin 

Accepted to ECCV 2024  

Paper  Video  Data 

Updates

• [07/2024] Dataset Omni6D released!
• [09/2024] Dataset Omni6D-xl and Omni6D-Real released!

Usage

Download the dataset

• Sign up here.
• Install OpenDataLab's CLI tools through pip install openxlab and update it through pip install -U openxlab.
• View and download the dataset from the command line:

openxlab login                                                  # Login, input AK/SK
openxlab dataset info --dataset-repo kszpxxzmcwww/Omni6D        # View dataset info
openxlab dataset ls --dataset-repo kszpxxzmcwww/Omni6D	        # View a list of dataset files
openxlab dataset get --dataset-repo kszpxxzmcwww/Omni6D         # Download the whole dataset (the compressed files require approximately 388.9GB of storage)

If you are experiencing an error of 401: {"msg":"login required"} with your own AKSK, please use the following AKSK:

AK: bmyqk5wpbaxl6x1vkzq9
SK: nl7kq9palyr6j3pwxolden7ezq4dwjmbgdm81yeo

You can check out the full folder structure on the website above and download a certain portion of the data by specifying the path. For example:

openxlab dataset download --dataset-repo kszpxxzmcwww/Omni6D \
                          --source-path /raw/Omni6D_V1/CAMERA/test \
                          --target-path <your-target-path>

For more information, please refer to the documentation.

Dataset format

Omni6D(-xl) // Omni6D(-xl) Dataset
├── info
│   ├── sym_info.csv 
│   │       // rotational invariance of each mesh
│   ├── <train/val/test/test_unseen>_list.txt
│   │       // valid data list of each split
├── shape_data               
│   ├── camera_<train/val/test/test_unseen>.pkl
│   │       // sampled 1024 points from each mesh, 
│   │       // normalized to NOCS
│   ├── ShapeNetCore_<2048/4096>.h5
│   ├── ShapeNetCore_unseen_<2048/4096>.h5
│   │       // sampled 2048/4096 points from each mesh
│   │       // normalized to NOCS
│   ├── mean_embedding.npy
│   │       // mean embedding of each category
│   ├── mean_points_emb.npy
│   │       // mean shape of each category
├── CAMERA         
│   ├── <train/val/test/test_unseen>
│   │   ├── <scene_id>
│   │   │   ├── <render_id>_color.png
│   │   │   │       // RGB image
│   │   │   ├── <render_id>_depth.png
│   │   │   │       // depth map
│   │   │   ├── <render_id>_coord.png
│   │   │   │       // NOCS mapping
│   │   │   ├── <render_id>_mask.png
│   │   │   │       // instance mask
│   │   │   ├── <render_id>_meta.txt
│   │   │   │       // instance name
│   │   │   ├── <render_id>_label.pkl
│   │   │   │       // ground truth annotations
Omni6D_Real // Omni6D_Real Dataset
│   ├── <scene_id>
│   │   ├── <capture_id>_color.png
│   │   │       // RGB image
│   │   ├── <capture_id>_depth.png
│   │   │       // depth map
│   │   ├── <capture_id>_mask.png
│   │   │       // instance mask
│   │   ├── <capture_id>_mask.png
│   │   │       // instance name
│   │   ├── <capture_id>_label.pkl
│   │   │       // ground truth annotations

Dataset construction

1. Install a suitable version of BlenderProc

git clone [email protected]:3DTopia/Omni6D.git
cd Omni6D
pip install blenderproc 

2. In dataset/dataset.py, update the following:

• Replace <path-to-output> with the path to your dataset.

• Replace <path-to-replica> with the path to your replica files.

  The scenes can be downloaded from Replica-Dataset.

• Replace <path-to-your-objectmesh> with the path to your object meshes.

  Our object meshes will not be released due to their large size and because they are not necessary for the task. However, you can find the corresponding point clouds in Omni6D(-xl)/shape_data/ShapeNetCore_<2048/4096>.h5, which contain 2048 or 4096 points sampled from each mesh, normalized to NOCS.

  For constructing your own dataset with real-scanned object meshes, you can make use of OmniObject3D, which includes a significant portion of our object meshes.

4. For a single test, run the following command:

blenderproc run dataset/dataset.py

4. For batch dataset generation, execute the batch script:

bash multi_run.sh

Benchmark

Category-level performance on Omni6D dataset.

Methods	Network	$IoU_{50}$	$IoU_{75}$	$5^\circ2cm$	$5^\circ5cm$	$10^\circ2cm$	$10^\circ5cm$	$5^\circ$	$10^\circ$	$2cm$	$5cm$
SPD	implicit	44.56	20.37	7.55	9.56	14.76	19.23	10.68	21.02	37.49	70.09
SGPA	implicit	36.34	14.44	4.78	6.84	10.13	15.03	8.49	17.73	25.57	59.18
DualPoseNet	hybrid	58.84	25.49	8.28	9.30	17.26	19.05	9.38	19.18	73.82	96.37
RBP-Pose	hybrid	35.92	4.66	0.37	0.60	0.53	0.80	0.75	0.96	39.73	83.55
GPV-Pose	explicit	15.28	0.26	0.10	0.70	0.14	0.96	2.25	2.96	5.31	33.70
HS-Pose	explicit	62.65	23.02	4.26	4.85	10.49	11.61	4.96	11.75	80.93	97.78

Category-level performance on Omni6D-xl dataset.

Methods	Network	$IoU_{50}$	$IoU_{75}$	$5^\circ2cm$	$5^\circ5cm$	$10^\circ2cm$	$10^\circ5cm$	$5^\circ$	$10^\circ$	$2cm$	$5cm$
SPD	implicit	42.28	16.73	3.42	6.47	6.45	12.16	10.62	17.71	21.05	54.14
SGPA	implicit	37.62	12.48	2.45	5.56	5.12	10.98	10.25	17.88	16.21	47.31
DualPoseNet	hybrid	59.15	24.26	5.92	7.24	10.74	13.00	7.70	13.72	65.05	87.44
RBP-Pose	hybrid	41.21	6.55	0.52	1.63	0.59	1.74	3.02	3.20	23.20	57.27
GPV-Pose	explicit	17.01	1.42	0.51	1.90	0.55	2.07	5.08	5.72	5.19	26.32
HS-Pose	explicit	62.92	26.50	4.98	6.26	8.81	10.73	6.65	11.39	71.68	89.38

License

The OmniObject3D dataset is released under the CC BY 4.0.

Reference

If you find our work helpful for your research, please consider giving a star ⭐ and citation 📝

@misc{zhang2024omni6dlargevocabulary3dobject,
      title={Omni6D: Large-Vocabulary 3D Object Dataset for Category-Level 6D Object Pose Estimation}, 
      author={Mengchen Zhang and Tong Wu and Tai Wang and Tengfei Wang and Ziwei Liu and Dahua Lin},
      year={2024},
      eprint={2409.18261},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2409.18261}, 
}