An unofficial Pytorch implementation of PV-RCNN: Point-Voxel Feature Set Abstraction for 3D Object Detection.
- Add database sampling augmentation (see augmentation.py for details).
- Add fast rotated nms on gpu for target assignment and inference (from detectron2).
- Code refactor and bug fixes.
- Simple inference (require only numpy array of raw points).
- Clean, testable codebase that's easy to debug.
- General 3D detection library (easy to extend to new models).
- Reproduce results of paper.
- This repo is under active development.
- I will post a pretrained model when codebase stabilizes and results are good.
- I will add more detailed training and inference instructions.
- I will add description of codebase.
See inference.py.
See install.md and please ask if you have any questions. I will supply a Docker build soon.
If you find this work helpful in your research, please consider starring this repo and citing:
@article{pvrcnnpytorch,
author={Jacob Hultman},
title={PV-RCNN PyTorch},
journal={https://github.com/jhultman/PV-RCNN},
year={2020}
}
and the original PV-RCNN paper (note I am not an author of this paper):
@article{shi2019pv,
author={Shi, Shaoshuai and Guo, Chaoxu and Jiang, Li and Wang, Zhe and Shi, Jianping and Wang, Xiaogang and Li, Hongsheng},
title={PV-RCNN: Point-Voxel Feature Set Abstraction for 3D Object Detection},
journal={arXiv preprint arXiv:1912.13192},
year={2019}
}
Contributions are welcome. Please post an issue if you find any bugs.
Please see license.md. Note that the code in pvrcnn/ops is largely from detectron2 and hence is subject to the Apache license. Thank you to the authors of PV-RCNN for their research.