# PETRv2: A Unified Framework for 3D Perception from Multi-Camera Images

**Authors:** Yingfei Liu, Junjie Yan, Fan Jia, Shuailin Li, Qi Gao, Tiancai Wang, Xiangyu Zhang, Jian Sun  
**arXiv:** [2206.01256](https://arxiv.org/abs/2206.01256)  
**Date:** 2022-06-02  
**Summary page:** https://www.summarizepaper.com/en/paper/2206.01256v1/

## Key points

- PETRv2 is a unified framework for 3D perception from multi-view images
- It explores the effectiveness of temporal modeling to enhance 3D object detection
- The framework extends the 3D position embedding (3D PE) to incorporate temporal modeling
- A feature-guided position encoder is introduced to improve adaptability to different datasets
- PETRv2 addresses high-quality Bird's Eye View (BEV) segmentation through segmentation queries
- The framework demonstrates state-of-the-art performance in both 3D object detection and BEV segmentation tasks
- Robustness analysis is conducted on the PETR framework to validate its reliability

## Summary

In their paper titled "PETRv2: A Unified Framework for 3D Perception from Multi-Camera Images," authors Yingfei Liu, Junjie Yan, Fan Jia, Shuailin Li, Qi Gao, Tiancai Wang, Xiangyu Zhang, and Jian Sun propose PETRv2 as a unified framework for 3D perception from multi-view images. Building upon the previous PETR framework, PETRv2 explores the effectiveness of temporal modeling by leveraging the temporal information from previous frames to enhance 3D object detection. The authors extend the 3D position embedding (3D PE) in PETR to incorporate temporal modeling. The 3D PE achieves temporal alignment by aligning the object positions across different frames. To improve its adaptability to different datasets, a feature-guided position encoder is introduced. <PETRv2 Framework>

PETRv2 also addresses high-quality Bird's Eye View (BEV) segmentation by introducing a simple yet effective solution through the addition of segmentation queries. Each segmentation query is responsible for segmenting a specific patch of the BEV map. The proposed framework demonstrates state-of-the-art performance in both 3D object detection and BEV segmentation tasks. <Temporal Modeling>

<Bird's Eye View Segmentation>

Additionally, detailed robustness analysis is conducted on the PETR framework to validate its reliability. Overall, this paper presents PETRv2 as a comprehensive and unified framework for 3D perception from multi-camera images. The proposed enhancements in temporal modeling and BEV segmentation contribute to improved performance in various perception tasks.