Focal Sparse Convolutional Networks for 3D Object Detection (CVPR 2022, Oral)

This is the official implementation of Focals Conv (CVPR 2022), a new sparse convolution design for 3D object detection (feasible for both lidar-only and multi-modal settings). For more details, please refer to:

Focal Sparse Convolutional Networks for 3D Object Detection [Paper]
Yukang Chen, Yanwei Li, Xiangyu Zhang, Jian Sun, Jiaya Jia

Experimental results

KITTI dataset

	Car@R11	Car@R40	download
PV-RCNN + Focals Conv	83.91	85.20	Google \| Baidu (key: m15b)
PV-RCNN + Focals Conv (multimodal)	84.58	85.34	Google \| Baidu (key: ie6n)
Voxel R-CNN (Car) + Focals Conv (multimodal)	85.68	86.00	Google \| Baidu (key: tnw9)

nuScenes dataset

	mAP	NDS	download
CenterPoint + Focals Conv (multi-modal)	63.86	69.41	Google \| Baidu (key: 01jh)
CenterPoint + Focals Conv (multi-modal) - 1/4 data	62.15	67.45	Google \| Baidu (key: 6qsc)

Visualization of voxel distribution of Focals Conv on KITTI val dataset:

Getting Started

Installation

a. Clone this repository

https://github.com/dvlab-research/FocalsConv && cd FocalsConv

b. Install the environment

Following the install documents for OpenPCdet and CenterPoint codebases respectively, based on your preference.

*spconv 2.x is highly recommended instead of spconv 1.x version.

c. Prepare the datasets.

Download and organize the official KITTI and Waymo following the document in OpenPCdet, and nuScenes from the CenterPoint codebase.

*Note that for nuScenes dataset, we include image-level gt-sampling (copy-paste) to train the multi-modal setting. Please download this dbinfos_train_10sweeps_withvelo.pkl to replace the original one. (Google | Baidu (key:b466))

*Note that for nuScenes dataset, we conduct ablation studies on a 1/4 data training split. Please download infos_train_mini_1_4_10sweeps_withvelo_filter_True.pkl if you needed for training. (Google | Baidu (key:769e))

Evaluation

We provide the trained weight file so you can just run with that. You can also use the model you trained.

For models in OpenPCdet,

NUM_GPUS=8
cd tools 
bash scripts/dist_test.sh ${NUM_GPUS} --cfg_file cfgs/kitti_models/voxel_rcnn_car_focal_multimodal.yaml --ckpt path/to/voxelrcnn_focal_multimodal.pth

bash scripts/dist_test.sh ${NUM_GPUS} --cfg_file cfgs/kitti_models/pv_rcnn_focal_multimodal.yaml --ckpt ../pvrcnn_focal_multimodal.pth

bash scripts/dist_test.sh ${NUM_GPUS} --cfg_file cfgs/kitti_models/pv_rcnn_focal_lidar.yaml --ckpt path/to/pvrcnn_focal_lidar.pth

For models in CenterPoint,

CONFIG="nusc_centerpoint_voxelnet_0075voxel_fix_bn_z_focal_multimodal"
python -m torch.distributed.launch --nproc_per_node=${NUM_GPUS} ./tools/dist_test.py configs/nusc/voxelnet/$CONFIG.py --work_dir ./work_dirs/$CONFIG --checkpoint centerpoint_focal_multimodal.pth

Training

For configures in OpenPCdet,

bash scripts/dist_train.sh ${NUM_GPUS} --cfg_file cfgs/kitti_models/CONFIG.yaml

For configures in CenterPoint,

python -m torch.distributed.launch --nproc_per_node=${NUM_GPUS} ./tools/train.py configs/nusc/voxelnet/$CONFIG.py --work_dir ./work_dirs/CONFIG

Note that we use 8 GPUs to train OpenPCdet models and 4 GPUs to train CenterPoint models.

TODO List

- Config files and trained models on the overall Waymo dataset.
- Config files and scripts for the test augs (double-flip and rotation) in nuScenes test submission.
- Results and models of Focals Conv Networks on 3D Segmentation datasets.

Citation

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

@inproceedings{focalsconv-chen,
  title={Focal Sparse Convolutional Networks for 3D Object Detection},
  author={Chen, Yukang and Li, Yanwei and Zhang, Xiangyu and Sun, Jian and Jia, Jiaya},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  year={2022}
}

Acknowledgement

This work is built upon the OpenPCDet and CenterPoint. Please refer to the official github repositories, OpenPCDet and CenterPoint for more information.
This README follows the style of IA-SSD.

License

This project is released under the Apache 2.0 license.

ShaoTengLiu/FocalsConv