BEVFusion: A Simple and Robust LiDAR-Camera Fusion Framework.

This paper focuses on LiDAR-camera fusion for 3D object detection. If you find this project useful, please cite:

@inproceedings{liang2022bevfusion,
  title={{BEVFusion: A Simple and Robust LiDAR-Camera Fusion Framework}},
  author={Tingting Liang, Hongwei Xie, Kaicheng Yu, Zhongyu Xia, Zhiwei Lin, Yongtao Wang, Tao Tang, Bing Wang and Zhi Tang},
  booktitle = {Neural Information Processing Systems (NeurIPS)},
  year={2022}
}

Introduction

Fusing the camera and LiDAR information has become a de-facto standard for 3D object detection tasks. Current methods rely on point clouds from the LiDAR sensor as queries to leverage the feature from the image space. However, people discover that this underlying assumption makes the current fusion framework infeasible to produce any prediction when there is a LiDAR malfunction, regardless of minor or major. This fundamentally limits the deployment capability to realistic autonomous driving scenarios. In contrast, we propose a surprisingly simple yet novel fusion framework, dubbed BEVFusion, whose camera stream does not depend on the input of LiDAR data, thus addressing the downside of previous methods. We empirically show that our framework surpasses the state-of-the-art methods under the normal training settings. Under the robustness training settings that simulate various LiDAR malfunctions, our framework significantly surpasses the state-of-the-art methods by 15.7% to 28.9% mAP. To the best of our knowledge, we are the first to handle realistic LiDAR malfunction and can be deployed to realistic scenarios without any post-processing procedure.

Main Results

nuScenes detection test

Model	Head	3DBackbone	2DBackbone	mAP	NDS	Link
BEVFusion	TransFusion-L	VoxelNet	Dual-Swin-T	69.2	71.8	Detection
BEVFusion*	TransFusion-L	VoxelNet	Dual-Swin-T	71.3	73.3	Leadboard

nuScenes detection validation

Model	Head	3DBackbone	2DBackbone	mAP	NDS	Model
BEVFusion	PointPillars	-	Dual-Swin-T	22.9	31.1	Model
BEVFusion	PointPillars	PointPillars	-	35.1	49.8	Model
BEVFusion	PointPillars	PointPillars	Dual-Swin-T	53.5	60.4	Model
BEVFusion	CenterPoint	-	Dual-Swin-T	27.1	32.1	-
BEVFusion	CenterPoint	VoxelNet	-	57.1	65.4	-
BEVFusion	CenterPoint	VoxelNet	Dual-Swin-T	64.2	68.0	-
BEVFusion	TransFusion-L	-	Dual-Swin-T	22.7	26.1	-
BEVFusion	TransFusion-L	VoxelNet	-	64.9	69.9	-
BEVFusion	TransFusion-L	VoxelNet	Dual-Swin-T	67.9	71.0	-
BEVFusion*	TransFusion-L	VoxelNet	Dual-Swin-T	69.6	72.1	Model

*: These methods exploit BEV-space data augmentation during training.

Acknowledgements of BEV-space augmentation

We conduct the BEV-space augmentation proposed by BEVDet and our implementation of GlobalRotScaleTransBEV and RandomFlip3DBEV followed BEVFusion-mit. You can also refer BEVDet, DEVDepth for more implementation versions.

Moreover, there are also image-sapce augmentations you can explore in LSS, BEVDet, BEVFusion-mit, DEVDepth.

Thanks for these attractive works and their great open-sourced code bases.

nuScenes detection validation against LiDAR Malfunctions

Infos for random box dropping in validation set are in drop_foreground.json, with LiDAR file name: dropping box (True) or not (False).

Model	Limited FOV	Objects failure	Head	3DBackbone	2DBackbone	mAP	NDS	Model
BEVFusion	(-π/3,π/3)	False	PointPillars	PointPillars	Dual-Swin-T	33.5	42.1	Model
BEVFusion	(-π/2,π/2)	False	PointPillars	PointPillars	Dual-Swin-T	36.8	45.8	Model
BEVFusion	-	True	PointPillars	PointPillars	Dual-Swin-T	41.6	51.9	Model
BEVFusion	(-π/3,π/3)	False	CenterPoint	VoxelNet	Dual-Swin-T	40.9	49.9	-
BEVFusion	(-π/2,π/2)	False	CenterPoint	VoxelNet	Dual-Swin-T	45.5	54.9	-
BEVFusion	-	True	CenterPoint	VoxelNet	Dual-Swin-T	54.0	61.6	-
BEVFusion	(-π/3,π/3)	False	TransFusion-L	VoxelNet	Dual-Swin-T	41.5	50.8	-
BEVFusion	(-π/2,π/2)	False	TransFusion-L	VoxelNet	Dual-Swin-T	46.4	55.8	-
BEVFusion	-	True	TransFusion-L	VoxelNet	Dual-Swin-T	50.3	57.6	-

nuImage detection validation

Model	2DBackbone	bbox_mAP	segm_mAP	Model
Mask R-CNN	Dual-Swin-T	56.0	46.1	Model

Use BEVFusion

Installation

Please refer to getting_started.md for installation of mmdet3d.

Recommended environments:

python==3.8.3
mmdet==2.11.0 (please install mmdet in mmdetection-2.11.0)
mmcv==1.4.0
mmdet3d==0.11.0
numpy==1.19.2
torch==1.7.0
torchvision==0.8.0

Benchmark Evaluation and Training

Please refer to data_preparation.md to prepare the data. Then follow the instruction there to train our model. All detection configurations are included in configs.

# training example for bevfusion-pointpillar 
# train nuimage for camera stream backbone and neck.
./tools/dist_train.sh configs/bevfusion/cam_stream/mask_rcnn_dbswin-t_fpn_3x_nuim_cocopre.py 8
# first train camera stream
./tools/dist_train.sh configs/bevfusion/cam_stream/bevf_pp_4x8_2x_nusc_cam.py 8
# then train LiDAR stream
./tools/dist_train.sh configs/bevfusion/lidar_stream/hv_pointpillars_secfpn_sbn-all_4x8_2x_nus-3d.py 8
# then train BEVFusion
./tools/dist_train.sh configs/bevfusion/bevf_pp_2x8_1x_nusc.py 8

### evaluation example for bevfusion-pointpillar
./tools/dist_test.sh configs/bevfusion/bevf_pp_2x8_1x_nusc.py ./work_dirs/bevfusion_pp.pth 8 --eval bbox

### training example for sota Results with BEV_augs
git checkout dev_aug
# first train camera stream
./tools/dist_train.sh configs/bevfusion/cam_stream/bevf_tf_4x8_20e_nusc_cam_lr.py 8
# then train LiDAR stream
./tools/dist_train.sh configs/bevfusion/lidar_stream/transfusion_nusc_voxel_L.py 8
# then train BEVFusion
./tools/dist_train.sh configs/bevfusion/bevf_tf_4x8_10e_nusc_aug.py 8

Acknowlegement

We sincerely thank the authors of mmdetection3d, TransFusion, LSS, CenterPoint for open sourcing their methods.

SxJyJay/BEVFusion