Introduction

This repository is an official implementation of RecurrentBEV. It is built based on MMDetection3D.

Main Results

NuScenes Val Set

Backbone	Img Size	Pretrain	NDS	mAP	Config	Download
Res50	256x704	ImageNet	54.9	44.5	config	model
Res101	512x1408	ImageNet	59.9	50.9	config	-
Res101	512x1408	NuImages	61.2	52.8	config	model

NuScenes Test Set

Backbone	Img Size	Pretrain	NDS	mAP	Config	Download
V2-99	640x1600	DD3D	65.1	57.3	config	model
ConvNeXt-B	640x1600	COCO	65.1	57.4	config	-

Inference Speed

The below table shows end-to-end FPS (Frames Per Second) of RecurrentBEV measured with a single RTX-3090.

Method	Pytorch-FP32	TensorRT-FP32	TensorRT-FP16	TensorRT-INT8
RecurrentBEV	25.6	46.3	129.3	234.8
StreamPETR	26.7	53.9	134.6	167.4

Getting Started

Please follow our documentation step by step. If you like our work, please recommend it to your colleagues and friends.

1. Environment Setup.

2. Data Preparation.

3. Training and Inference.

4. Visualization.

5. Deployment.

Features List

• RecurretBEV code
• Visualization
• Convert to TRT model
• TensorRT inference

Acknowledgements

We thank these great works and open-source codebases:

• MMDetection3D

• BEVDet

• SOLOFusion

• StreamPETR

Citation

If you find RecurrentBEV is useful in your research or applications, please consider giving us a star 🌟 and citing it by the following BibTeX entry.