Efficient Deformable Modeling Network for Multi-View 3D Object Detection

Publication: Efficient Deformable Modeling Network for Multi-view 3D Object Detection, Neural Computing and Applications (Springer Nature), vol. 38, article number 122 (https://doi.org/10.1007/s00521-026-11881-y), Feb. 14, 2026

Getting Started

Our implementation is based on StreamPETR. Please follow Environment Setup and Data Preparation step by step.

Train & Inference

Train

tools/dist_train.sh projects/configs/RepDETR4D/repdetr4d_res50_706_bs16_seq_60e.py 4 --work-dir work_dirs/RepDETR4D/

Evaluation

tools/dist_test.sh projects/configs/RepDETR4D/repdetr4d_res50_706_bs16_seq_60e.py work_dirs/RepDETR4D/latest.pth 4 --eval bbox

Results on NuScenes Val Set.

Model	Setting	Pretrain	Lr Schd	NDS	mAP	Config	Weight
StreamPETR	R18	ImageNet	60ep	48.4	36.2	-	-
RepPETR4D	R18	ImageNet	60ep	50.0	37.8	config	weight
StreamPETR	R50	NuImg	60ep	54.5	44.9	-
RepPETR4D	R50	NuImg	60ep	55.1	45.5	config	weight

Results on NuScenes Test Set.

Model	Setting	Pretrain	NDS	mAP
StreamPETR	R50	NuImg	56.3	46.0
RepPETR4D	R50	NuImg	56.7	46.9

Acknowledgements

We thank these great works and open-source codebases:

• 3D Detection. MMDetection3d, DETR3D, PETR, BEVFormer, SOLOFusion, Sparse4D, StreamPETR.