Multiview 3D Object Detection Based on Improved DETR3D

Yuhan Zhang; Miaohua Huang; Gengyao Chen; Yanzhou Li; Yiming Wu

doi:10.3788/LOP240912

Journals >Laser & Optoelectronics Progress >Volume 62 >Issue 2 >Page 0212001 > Article

Laser & Optoelectronics Progress
Vol. 62, Issue 2, 0212001 (2025)

Multiview 3D Object Detection Based on Improved DETR3D

Yuhan Zhang^1,2,3, Miaohua Huang^1,2,3,*, Gengyao Chen^1,2,3, Yanzhou Li^1,2,3, and Yiming Wu^1,2,3

Author Affiliations

¹Hubei Key Laboratory of Advanced Technology for Automotive Components, Wuhan University of Technology, Wuhan 430070, Hubei , China

²Hubei Collaborative Innovation Center for Automotive Components Technology, Wuhan University of Technology, Wuhan 430070, Hubei , China

³Hubei Research Center for New Energy & Intelligent Connected Vehicle, Wuhan University of Technology, Wuhan 430070, Hubei , China

show less

DOI: 10.3788/LOP240912 Cite this Article Set citation alerts

Yuhan Zhang, Miaohua Huang, Gengyao Chen, Yanzhou Li, Yiming Wu. Multiview 3D Object Detection Based on Improved DETR3D[J]. Laser & Optoelectronics Progress, 2025, 62(2): 0212001 Copy Citation Text

show less

Fig. 1. Algorithm overall framework diagram

Download full size

Fig. 2. 3D position encoding process

Download full size

Fig. 3. Perceptual ability of queries for other queries from different distances

Download full size

Fig. 4. Temporal feature sampling process

Download full size

Fig. 5. Visualization of baseline model predictions

Download full size

Fig. 6. Visualization of improved model predictions

Download full size

Algorithm	Modality	NDS	mAP	mATE	mASE	mAOE	mAVE	mAAE	FPS
DETR3D	C	0.434	0.349	0.716	0.268	0.379	0.842	0.200	3.7
BEVDet	C	0.417	0.349	0.637	0.269	0.490	0.914	0.268	16.7
PETR	C	0.441	0.366	0.717	0.267	0.412	0.834	0.190	5.7
BEVFormer	C	0.517	0.416	0.673	0.274	0.372	0.394	0.198	3.0
BEVDepth	C	0.535	0.412	0.565	0.266	0.358	0.331	0.190	4.5
Proposed algorithm	C	0.508	0.467	0.564	0.445	0.518	0.443	0.289	15.6

Table 1. Comparison of results of different algorithms on the nuScenes validation set

Algorithm	Modality	NDS	mAP	mATE	mASE	mAOE	mAVE	mAAE
PointPillar	L	0.453	0.305
CenterPoint	L	0.655	0.580
DETR3D	C	0.479	0.412	0.641	0.255	0.394	0.845	0.133
BEVDet	C	0.488	0.424	0.524	0.242	0.373	0.950	0.148
PETR	C	0.504	0.441	0.593	0.249	0.383	0.808	0.132
BEVFormer	C	0.569	0.481	0.582	0.256	0.375	0.378	0.126
BEVDepth	C	0.609	0.520	0.445	0.243	0.352	0.347	0.127
Proposed algorithm	C	0.614	0.514	0.462	0.245	0.343	0.266	0.114

Table 2. Comparison of results of different algorithms on the nuScenes test set

Module	Setting	NDS	mAP
Query formulation	3D reference point	0.479	0.435
Query formulation	Ours	0.489	0.443
3DPE	None	0.474	0.431
3DPE	Ours	0.489	0.443
MSSA	MHSA	0.467	0.403
MSSA	Ours	0.489	0.443
TFFS	N=4/8/16	0.459/0.475/0.489	0.407/0.422/0.443
TFFS	T=1/4/8	0.401/0.462/0.489	0.358/0.415/0.443

Table 3. Impact of different modules on model performance

Yuhan Zhang, Miaohua Huang, Gengyao Chen, Yanzhou Li, Yiming Wu. Multiview 3D Object Detection Based on Improved DETR3D[J]. Laser & Optoelectronics Progress, 2025, 62(2): 0212001

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information

微信扫一扫：分享

微信扫一扫：分享