Research map / Detection and segmentation / 3D detection and driving
BEVFormer v2: Adapting Modern Image Backbones to Bird's-Eye-View Recognition via Perspective Supervision
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 perspective supervision to BEV supervision (ablation) | BEVFormer v2 with ResNet-101 · nuScenes val | NDS: 0.426 -> 0.451 | Helped |
| Replace the perspective proposal head with a second BEV head (ablation) | two-stage BEV detector with ResNet-101 · nuScenes val | NDS: 0.451 -> 0.428 | Hurt |
| Add perspective supervision (ablation) | BEV detector across ResNet-50, DLA-34, ResNet-101, VoVNet-99, and InternImage-B · nuScenes val | NDS: 0.400 -> 0.428; 0.403 -> 0.435; 0.426 -> 0.451; 0.441 -> 0.467; 0.455 -> 0.485 | Helped |
| Add perspective supervision (ablation) | ResNet-50 BEV detector · nuScenes val | NDS: 0.379 -> 0.414 (24 epochs); 0.410 -> 0.428 (72 epochs) | Helped |
| Extend BEV-only training from 24 to 72 epochs (ablation) | ResNet-50 BEV Only · nuScenes val | NDS: 0.379 -> 0.410 | Helped |
| Use DD3D instead of DETR3D for the perspective head (ablation) | BEVFormer v2 with Deformable DETR and ResNet-50 · nuScenes val | NDS: 0.409 -> 0.428 | Helped |
| Replace Deformable DETR with Group DETR for the BEV head (ablation) | BEVFormer v2 with DD3D and ResNet-50 · nuScenes val | NDS: 0.428 -> 0.445 | Helped |
| Add image-level data augmentation (ablation) | ResNet-50 BEVFormer v2 with temporal information · nuScenes val | NDS: 0.478 -> 0.489 | Helped |
| Use a longer temporal interval (ablation) | ResNet-50 BEVFormer v2 with image-level augmentation · nuScenes val | NDS: 0.489 -> 0.498 | Helped |
| Add a bidirectional temporal encoder (ablation) | ResNet-50 BEVFormer v2 with long temporal interval · nuScenes val | NDS: 0.498 -> 0.529 | Helped |
| Remove perspective supervision while retaining the other listed components (ablation) | ResNet-50 BEVFormer v2 with IDA, long interval, and bidirectional temporal encoder · nuScenes val | NDS: 0.529 -> 0.507 | Hurt |
| Use BEV supervision only instead of combining perspective and BEV supervision (ablation) | BEV detector with ResNet-101 · nuScenes val | mAP: 0.355 -> 0.374 | Hurt |
Closest papers by meaning
- A Black-Box Evaluation Framework for Semantic Robustness in Bird’s Eye View Detection (AAAI 2025 · similarity 0.88)
- CLIP-BEVFormer: Enhancing Multi-View Image-Based BEV Detector with Ground Truth Flow (CVPR 2024 · similarity 0.87)
- Instance-Aware Multi-Camera 3D Object Detection with Structural Priors Mining and Self-Boosting Learning (AAAI 2024 · similarity 0.87)
- BEVNeXt: Reviving Dense BEV Frameworks for 3D Object Detection (CVPR 2024 · similarity 0.87)
- Improving Bird's Eye View Semantic Segmentation by Task Decomposition (CVPR 2024 · similarity 0.87)
- GeoBEV: Learning Geometric BEV Representation for Multi-view 3D Object Detection (AAAI 2025 · similarity 0.86)
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.