Research map / Detection and segmentation / 3D detection and driving
PETRv2: A Unified Framework for 3D Perception from Multi-Camera Images
Yingfei Liu, Junjie Yan, Fan Jia, Shuailin Li, Aqi Gao, Tiancai Wang, Xiangyu Zhang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add temporal frame to PETR without coordinate alignment (ablation) | PETR · nuScenes val (VoV-99) | NDS: 0.434 -> 0.461 | Improved |
| add 3D coordinates alignment (CA) of previous frame (ablation) | PETRv2 · nuScenes val (VoV-99) | NDS: 0.461 -> 0.482 | Improved |
| add CA on mAVE (ablation) | PETRv2 · nuScenes val (VoV-99) | mAVE: 0.605 -> 0.429 | Improved |
| add feature-guided position encoder (FPE) (ablation) | PETRv2 · nuScenes val (VoV-99) | NDS: 0.482 -> 0.496 | Improved |
| add FPE to single-frame PETR (ablation) | PETR · nuScenes val (VoV-99) | NDS: 0.434 -> 0.449 | Improved |
| add FPE for robustness to extrinsics noise (ablation) | PETR · nuScenes val, Rmax=8 | mAP drop: 6.33% -> 4.68% mAP drop | Improved |
| add temporal multi-frame extension for robustness to extrinsics noise (ablation) | PETR+FPE · nuScenes val, extrinsics noise | mAP drop: 4.68% -> 4.12% mAP drop (Rmax=8) | No clear effect |
| use 400 anchor points instead of 10 per lane (ablation) | PETRv2-V · OpenLane | F-score (%): 57.8 -> 61.2 | Improved |
| stronger pretrained VoV-99 backbone instead of EfficientNet | PETRv2 · OpenLane | F-score (%): 51.9 -> 57.8 | Improved |
| camera time delay of one frame interval at inference | PETRv2 · nuScenes val | mAP (%): -3.19 mAP, -8.4 NDS (delay 0.083s) | Hurt |
| drop the back camera at test time | PETRv2 · nuScenes val | mAP (%): -13.19% mAP | Hurt |
Later papers that compared with this method
14 later accepted papers report PETRv2 as a baseline in their results tables; 12 of them report a better number on most of the results they share with it.
- CycleBEV: Regularizing View Transformation Networks via View Cycle Consistency for Bird's-Eye-View Semantic Segmentation (CVPR 2026 · better on 3 of 3 shared results)
- SemLT3D: Semantic-Guided Expert Distillation for Camera-only Long-Tailed 3D Object Detection (CVPR 2026 · better on 2 of 2 shared results)
- RepVF: A Unified Vector Fields Representation for Multi-task 3D Perception (ECCV 2024 · better on 1 of 1 shared result)
- Learning High-resolution Vector Representation from Multi-Camera Images for 3D Object Detection (ECCV 2024 · better on 1 of 1 shared result)
- HENet: Hybrid Encoding for End-to-end Multi-task 3D Perception from Multi-view Cameras (ECCV 2024 · better on 0 of 1 shared result)
- Object as Query: Lifting Any 2D Object Detector to 3D Detection (ICCV 2023 · better on 6 of 6 shared results)
- HotBEV: Hardware-oriented Transformer-based Multi-View 3D Detector for BEV Perception (NeurIPS 2023 · better on 3 of 3 shared results)
- FrustumFormer: Adaptive Instance-Aware Resampling for Multi-View 3D Detection (CVPR 2023 · better on 2 of 2 shared results)
6 more not listed.
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.
- BEVFormer — BEVFormer: Learning Bird’s-Eye-View Representation from Multi-Camera Images via Spatiotemporal Transformers (ECCV 2022) · better on 4 of 4 shared results
- PETR — PETR: Position Embedding Transformation for Multi-View 3D Object Detection (ECCV 2022) · better on 3 of 3 shared results
- PersFormer — PersFormer: 3D Lane Detection via Perspective Transformer and the OpenLane Benchmark (ECCV 2022) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- PETR: Position Embedding Transformation for Multi-View 3D Object Detection (ECCV 2022 · similarity 0.92)
- FQ-PETR: Fully Quantized Position Embedding Transformation for Multi-View 3D Object Detection (AAAI 2026 · similarity 0.88)
- STUR3D: Spatio-Temporal Unified Representation Learning for 3D Object Detection (CVPR 2026 · similarity 0.87)
- PE3R: Perception-Efficient 3D Reconstruction (CVPR 2026 · similarity 0.86)
- Towards Generalizable Multi-Camera 3D Object Detection via Perspective Rendering (AAAI 2025 · similarity 0.85)
- Perception Encoder: The best visual embeddings are not at the output of the network (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
