Research map / Detection and segmentation / 3D detection and driving
BEVFormer: Learning Bird’s-Eye-View Representation From LiDAR-Camera via Spatiotemporal Transformers
Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lü, Yu Qiao, Jifeng Dai
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 self-attention (ablation) | BEVFormer-S (R101) · nuScenes val | NDS: 0.448 -> 0.517 | Improved |
| add temporal self-attention (ablation) | BEVFormer-S (V2-99) · nuScenes test | NDS: 0.495 -> 0.569 | Improved |
| use deformable local attention instead of global attention in spatial cross-attention (ablation) | BEVFormer-S · nuScenes val | NDS: 0.404 -> 0.448 | Improved |
| sample local regions around reference points instead of reference points only (ablation) | BEVFormer-S · nuScenes val | NDS: 0.423 -> 0.448 | Improved |
| align history BEV features with ego-motion (ablation) | BEVFormer · nuScenes val | NDS: 0.510 -> 0.517 | Improved |
| randomly sample 4 frames from 5 during training (ablation) | BEVFormer · nuScenes val | NDS: 0.513 -> 0.517 | Improved |
| predict offsets and weights from both BEV queries and history BEV (ablation) | BEVFormer · nuScenes val | NDS: 0.513 -> 0.517 | Improved |
| increase training frames from 1 to 4 (ablation) | BEVFormer · nuScenes val | NDS: 0.448 -> 0.517 | Improved |
| increase training frames from 4 to 5 (ablation) | BEVFormer · nuScenes val | NDS: 0.517 -> 0.517 | No clear effect |
| use multi-scale view features (ablation) | BEVFormer · nuScenes val | NDS: 0.511 -> 0.517 | Improved |
| use 6 encoder layers instead of 1 (ablation) | BEVFormer · nuScenes val | NDS: 0.501 -> 0.517 | Improved |
| joint detection and segmentation training instead of segmentation only | BEVFormer · nuScenes val | Lane IoU: 25.7 -> 23.9 | Hurt |
Closest papers by meaning
- BEVFormer: Learning Bird’s-Eye-View Representation from Multi-Camera Images via Spatiotemporal Transformers (ECCV 2022 · similarity 0.98)
- BEV-MAE: Bird’s Eye View Masked Autoencoders for Point Cloud Pre-training in Autonomous Driving Scenarios (AAAI 2024 · similarity 0.90)
- Improving Bird's Eye View Semantic Segmentation by Task Decomposition (CVPR 2024 · similarity 0.90)
- M-BEV: Masked BEV Perception for Robust Autonomous Driving (AAAI 2024 · similarity 0.90)
- CALICO: Self-Supervised Camera-LiDAR Contrastive Pre-training for BEV Perception (ICLR 2024 · similarity 0.89)
- Map It Anywhere: Empowering BEV Map Prediction using Large-scale Public Datasets (NeurIPS 2024 · similarity 0.88)
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.
