/BEVDistill

[ICLR 2023] BEVDistill: Cross-Modal BEV Distillation for Multi-View 3D Object Detection

Primary LanguagePython

BEVDistill

Cross-Modal BEV Distillation for Multi-View 3D Object Detection


Zehui Chen, Zhenyu Li, Shiquan Zhang, Liangji Fang, Qinhong Jiang, Feng Zhao.
[Paper]

NOTE

This code is an initial release with BEVFormer-T only, we will release a full version in the future.

Run

# baseline
./tools/dist_train.sh projects/configs/video_bevformer/bevformer_128x128_r50_2x.py 8
# BEVDistill
./tools/dist_train_distill.sh projects/configs/bevdistill/bevdistill_r50_128x128_900query_bboxrweighttop100_bevgt_mocov2_merge_r1_2x.py 8

Performance

nuScenes Val set

Model config mAP NDS
BEVFormer-R50 35.2 42.3
BEVDistill-R50 38.6 45.7
BEVFormer-T-R50 link 38.3 48.8
BEVDistill-T-R50 link 40.7 51.5
BEVFormer-T-R101 link 40.5 50.6
BEVDistill-T-R101 link 41.7 52.4

Pretrain Checkpoints

Model Weight mAP NDS
ObjectDGCNN (Teacher) link 67.4 61.5
FCOS3D r50 (Student) link