EA-BEV: Edge-aware Bird' s-Eye-View Projector for 3D Object Detection
News
- 2023.4.6 create README.md
Main Result
nuScenes detection test
| Method | mAP | NDS |
|---|---|---|
| BEVFusion(Peking University) | 71.3 | 73.3 |
| +EA-BEV | 71.8 | 73.6 |
nuScenes detection validation
| Method | mAP | NDS | Latency(ms) |
|---|---|---|---|
| BEVDepth | 35.1 | 47.5 | 110.3 |
| +EA-BEV | 40.4 | 48.2 | 114.8 |
| BEVFusion(MIT) | 68.5 | 71.4 | 119.2 |
| +EA-BEV | 69.4 | 71.8 | 123.6 |
| BEVFusion(Peking University) | 69.6 | 72.1 | 190.3 |
| +EA-BEV | 70.3 | 72.6 | 194.9 |
nuScenes BEV map segmentation validation
| Method | Drivable | Ped.Cross. | Walkway | Stop Line | Carpark | Divider | Mean |
|---|---|---|---|---|---|---|---|
| BEVFusion(MIT) | 85.5 | 60.5 | 67.6 | 52.0 | 57.0 | 53.7 | 62.7 |
| +EA-BEV | 85.8 | 61.1 | 68.0 | 52.3 | 56.8 | 54.5 | 63.1 |
Visualization results
nuScenes 3D object detection
nuScenes map segmentation
Use EA-BEV
install and date preparation
For environment installation method, please refer to BEVFusion.
beachmark Evaluation and Training
# training example for EA-BEV
# first train camera stream
./tools/dist_train.sh configs/EA-BEV/cam_stream/eabev_tf_4x8_20e_nusc_cam_lr.py 8
# then train LiDAR stream (using fade strategy)
./tools/dist_train.sh configs/EA-BEV/lidar_stream/transfusion_nusc_voxel_L.py 8
# then train BEVFusion
./tools/dist_train.sh configs/EA-BEV/eabev_tf_4x8_10e_nusc_aug.py 8
### evaluation example for EA-BEV
./tools/dist_test.sh configs/EA-BEV/eabev_ft_4x8_10e_nusc_aug.py ./work_dirs/ea-bev_tf.pth 8 --eval bboxAcknowlegement
We sincerely thank the authors of BEVFusion(Peking University), BEVFusion(MIT), mmdetection3d, TransFusion for open sourcing their methods.