This is the official implementation of:

Fully Sparse 3D Object Detection and Embracing Single Stride 3D Object Detector with Sparse Transformer.

🔥 FSD Preview Release

• Code of SpConv-based FSD on Waymo is released. See ./configs/fsd/fsd_waymoD1_1x.py
• We provide the tools for processing Argoverse 2 dataset in ./tools/argo. We will release the instruction and configs of Argo2 model later.
• A very fast Waymo evaluation, see Usage section for detailed instructions. The whole evaluation process of FSD on Waymo costs less than 10min with 8 2080Ti GPUs.
• We cannot distribute model weights of FSD on Waymo due to the license. Users could contact us for the private model weights.
• Before using this repo, please install TorchEx and SpConv2 (SpConv 1.x is not supported).

NEWS

• [22-09-19] The code of FSD is released here.
• [22-09-15] 🔥 FSD is accepted at NeurIPS 2022.
• [22-06-06] Support SST with CenterHead, cosine similarity in attention, faster SSTInputLayer. See Usage for details.
• [22-03-02] 🔥 SST is accepted at CVPR 2022.
• Support Weighted NMS (CPU version) in RangeDet, improving performance of vehicle class by ~1 AP. See Usage section.
• We refactored the code to provide more clear function prototypes and a better understanding. See ./configs/sst_refactor
• Supported voxel-based region partition in ./configs/sst_refactor. Users can easily use voxel-based SST by modifying the recover_bev function in the backbone.
• Waymo Leaderboard results updated in SST_v1

PyTorch >= 1.9 is recommended for a better support of the checkpoint technique.

Our implementation is based on MMDetection3D, so just follow their getting_started and simply run the script: run.sh.

ATTENTION: It is highly recommended to check the data version if users generate data with the official MMDetection3D. MMDetection3D refactors its coordinate definition after v1.0. A hotfix is using our code to re-generate the waymo_dbinfo_train.pkl

• Copy tools/idx2timestamp.pkl and tools/idx2contextname.pkl to ./data/waymo/kitti_format/.
• Passing the argument --eval fast (See run.sh). This argument will directly convert network outputs to Waymo .bin format, which is much faster than the old way.
• Users could further build the multi-thread Waymo evaluation tool (link) for faster evaluation.

We only provide the single-stage model here, as for our two-stage models, please follow LiDAR-RCNN. It's also a good choice to apply other powerful second stage detectors to our single-stage SST.

We borrow Weighted NMS from RangeDet and observe ~1 AP improvement on our best Vehicle model. To use it, you are supposed to clone RangeDet, and simply run pip install -v -e . in its root directory. Then refer to config/sst/sst_waymoD5_1x_car_8heads_wnms.py to modify your config and enable Weight NMS. Note we only implement the CPU version for now, so it is relatively slow. Do NOT use it on 3-class models, which will lead to performance drop.

A basic config of SST with CenterHead: ./configs/sst_refactor/sst_waymoD5_1x_3class_centerhead.py, which has significant improvement in Vehicle class. To enable faster SSTInputLayer, clone https://github.com/Abyssaledge/TorchEx, and run pip install -v ..

Please refer to this page.

Please visit the website for detailed results: SST_v1

Note that we train the 3 classes together, so the performance above is a little bit lower than that reported in our paper.

Please consider citing our work as follows if it is helpful.

@inproceedings{fan2022embracing,
  title={{Embracing Single Stride 3D Object Detector with Sparse Transformer}},
  author={Fan, Lue and Pang, Ziqi and Zhang, Tianyuan and Wang, Yu-Xiong and Zhao, Hang and Wang, Feng and Wang, Naiyan and Zhang, Zhaoxiang},
  booktitle={CVPR},
  year={2022}
}

@article{fan2022fully,
  title={{Fully Sparse 3D Object Detection}},
  author={Fan, Lue and Wang, Feng and Wang, Naiyan and Zhang, Zhaoxiang},
  journal={arXiv preprint arXiv:2207.10035},
  year={2022}
}

This project is based on the following codebases.

• MMDetection3D
• LiDAR-RCNN

Thank the authors of CenterPoint for providing their detailed results.