Skip to content

Latest commit

 

History

History
100 lines (81 loc) · 3.63 KB

README.md

File metadata and controls

100 lines (81 loc) · 3.63 KB
Raw

OV9D: Open-Vocabulary Category-Level 9D Object Pose and Size Estimation

Junhao Cai1*, Yisheng He2*, Weihao Yuan2, Siyu Zhu3, Zilong Dong2, Liefeng Bo2, Qifeng Chen1

1The Hong Kong University of Science and Technology,   2Alibaba Group,   3Fudan University   
*Equal contribution

Project Page | Paper

Teaser

Environment

git clone https://github.com/caijunhao/ov9d.git
cd ov9d
conda env create --file env.yml
conda activate ov9d

Dataset

You can have access to the data by running the script

python download_data.py

If you want to download the object mesh models or the multi-object data, you can run the script with --models or --multi.

python download_data.py --models --multi

Split train and test data

python prepare_data.py

The structure of the dataset is organized as follow:

├── ov9d
│   | cid2oid.json  # mapping from class id to object id
│   | class_list.json  # name list of object category
│   | models_info.json  # model information in BOP format
│   | models_info_with_symmetry.json  # model information with symmetry axes
│   | name2oid.json  # mapping from object category to object id
│   | oid2cid.json  # mapping from object id to class id
│   | oo3d9dsingle_class_embeddings.json  # saved text features of object descriptions
│   | models
│     ├── obj_000001.ply
│     ├── ...
│   | models_eval
│     ├── obj_000001.ply
│     ├── ...
│   | oo3d9dsingle
│     ├── anise_001_xxx  # data folder for the object anise_001 organized in BOP format
│     ├── ...
│   | oo3d9dsingle
│     ├── oo3d-xxx  # data folder for the multiple objects organized in BOP format
│     ├── ...
│   | train  # folder containing softlinks of folders from oo3d9dsingle
│     ├── anise_001_xxx  # softlink of oo3d9dsingle/anise_001_xxx
│   | test  # folder containing softlinks of folders from oo3d9dsingle
│     ├── bowl  # folder of object instances with bowl category
│         ├── bowl_001_xxx  # softlink of oo3d9dsingle/bowl_001_xxx
│     ├── ...
│   | pretrained_model
│     ├── model.ckpt  # trained checkpoint

Training

./train.sh

Test

Here is an example of evaluation on object instances with bowl category using our training model.

./test.sh test/bowl ov9d/pretrained_model/model.ckpt

BibTex

If you find this project useful in your research, please cite:

@article{cai2024ov9d,
  title={Ov9d: Open-vocabulary category-level 9d object pose and size estimation},
  author={Cai, Junhao and He, Yisheng and Yuan, Weihao and Zhu, Siyu and Dong, Zilong and Bo, Liefeng and Chen, Qifeng},
  journal={arXiv preprint arXiv:2403.12396},
  year={2024}
}

Acknowledgements

This project is developed upon VPD, stable-diffusion, and NOCS. We thank the authors for open sourcing their great works!