Research map / 3D vision / 3D generation
UniFusion: Unified Multi-View Fusion Transformer for Spatial-Temporal Representation in Bird's-Eye-View
Zequn Qin, Jingyu Chen, Chao Chen, Xiaozhi Chen, Xi Li
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 fusion steps (0 to 1) (ablation) | UniFusion · nuScenes 100m x 100m | Road mIoU: 79.04 -> 79.48 | Improved |
| infer with 6 temporal steps while training with only 1 step (ablation) | UniFusion · nuScenes 100m x 100m | Road mIoU: 79.48 -> 81.12 | Improved |
| train with more fusion steps (1 vs 4) given 6 inference steps (ablation) | UniFusion · nuScenes 100m x 100m | Road mIoU: 81.12 -> 81.25 | No clear effect |
| extend temporal fusion up to 10 steps in unified fusion versus warp-based BEVFormer (ablation) | UniFusion · nuScenes | mIoU | Improved |
| add self-regression mechanism (12 layers) (ablation) | UniFusion · nuScenes 100m x 100m | Road mIoU: 81.13 -> 81.97 | Improved |
| add self-regression mechanism (6 layers) (ablation) | UniFusion · nuScenes 100m x 100m | Road mIoU: 80.42 -> 80.91 | Improved |
| increase Transformer layers from 6 to 12 (ablation) | UniFusion · nuScenes 100m x 100m | Lane mIoU: 24.99 -> 25.76 | Improved |
| use learned adaptive temporal fusion instead of equal-weighted averaging (6 steps) (ablation) | UniFusion · nuScenes 100m x 100m | Lane mIoU: 25.08 -> 25.76 | Improved |
| UniFusion ResNet50 versus BEVFormer with larger backbone and more queries | UniFusion · nuScenes 100m x 100m | Road mIoU: 80.1 -> 82.0 | Improved |
| switch from vanilla split to city-based split | UniFusion and baselines · nuScenes 160m x 100m easy | All mIoU: 40.9 -> 16.7 | Hurt |
| larger VoVNet99 backbone instead of Swin-tiny | UniFusion · nuScenes 60m x 30m | All mIoU: 53.6 -> 57.4 (vanilla); 26.5 -> 25.6 (city-based) | Mixed |
Closest papers by meaning
- Improving Bird's Eye View Semantic Segmentation by Task Decomposition (CVPR 2024 · similarity 0.87)
- IS-Fusion: Instance-Scene Collaborative Fusion for Multimodal 3D Object Detection (CVPR 2024 · similarity 0.87)
- MamBEV: Enabling State Space Models to Learn Birds-Eye-View Representations (ICLR 2025 · similarity 0.86)
- Toward Real-world BEV Perception: Depth Uncertainty Estimation via Gaussian Splatting (CVPR 2025 · similarity 0.86)
- MapExpert: Online HD Map Construction with Simple and Efficient Sparse Map Element Expert (AAAI 2025 · similarity 0.85)
- M-BEV: Masked BEV Perception for Robust Autonomous Driving (AAAI 2024 · similarity 0.85)
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.
