SimPB

[ECCV 2024] SimPB: A Single Model for 2D and 3D Object Detection from Multiple Cameras

Introduction

This repository is an official implementation of SimPB, which Simultaneously detects 2D objects in the Perspective view and 3D objects in the BEV space from multiple cameras.

Getting started

• Prepare Environment
• Prepare Dataset
• Training and Evaluation

Model Zoo

Results on NuScenes validation

method	backbone	pretrain	img size	mAP	NDS	config	ckpt	log
SimPB+	ResNet50	ImageNet	704x256	0.479	0.586	config	ckpt	log
SimPB+	ResNet50	nuImg	704x256	0.489	0.591	config	ckpt	log
SimPB	ResNet101	nuImg	1408x512	0.539	0.629

Note: SimPB+ is a modified architecture that introduces 2d denoise and removes the encoder. This slightly reduces the runtime while maintaining comparable performance compared to the released script.

Acknowledgement

Thanks to these excellent open-source works:

Sparse4Dv3, StreamPETR, SparseBEV, Far3D, MMDetection3D

Citation

 @article{simpb,
  title={SimPB: A Single Model for 2D and 3D Object Detection from Multiple Cameras},
  author={Yingqi Tang and Zhaotie Meng and Guoliang Chen and Erkang Cheng},
  journal={ECCV},
  year={2024}
}