Skip to content

Research platform for 3D object detection in PyTorch.

License

Notifications You must be signed in to change notification settings

jhultman/vision3d

Repository files navigation

Vision 3D

A clean, easy-to-use PyTorch library for lidar perception.

Project goals

  • Emphasis on clean code (no 1,000 LOC functions).
  • General 3D detection library (easy to extend to new models and datasets).

Status

  • This project is not under active development.
  • Implementation of SECOND is complete.
  • Implementation of PV-RCNN is partially completed.
  • 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.

Sample results on validation data (KITTI)

Sample result

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.