Vision 3D

A clean, easy-to-use PyTorch library for lidar perception. Currently supports SECOND detector.

Project goals

• Emphasis on simple codebase (no 1,000 LOC functions).
• General 3D detection library (easy to extend to new models and datasets).
• Hope to reproduce state-of-the-art results.

Status and plans

• At this time I do not have capacity to develop this project. Community support is welcomed.
• I hope this project can serve as a useful starting point for lidar perception research.
• Implementation of PV-RCNN is work-in-progress.
• These forks (one, two) have shown some promise in training on other datasets (NuScenes, and proprietary lidar data).

Usage

See inference.py and train.py. To train, need to first start a visdom server using command visdom to enable train loss monitoring. (Requires visdom python package to be installed).

Installation

See install.md.

Preliminary results

Citing

If you find this work helpful in your research, please consider starring this repo and citing:

@article{hultman2020vision3d,
  author={Jacob Hultman},
  title={vision3d},
  journal={https://github.com/jhultman/vision3d},
  year={2020}
}

Contributions

Contributions are welcome. Please post an issue if you find any bugs.

Acknowledgements and licensing

Please see license.md. Note that the code in vision3d/ops is largely from detectron2 and hence is subject to the Apache license.