Research map / Detection and segmentation / 3D detection and driving
MatrixVT: Efficient Multi-Camera to BEV Transformation for 3D Perception
Hongyu Zhou, Zheng Ge, Zeming Li, 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 Prime Extraction (compress image feature and depth) to BEVDepth | BEVDepth Res-50 · nuScenes val | view transformation memory: 734M -> 316M | Improved |
| add Prime Extraction to BEVDepth | BEVDepth Res-50 · nuScenes val | NDS: 0.414 -> 0.411 | No clear effect |
| add Prime Extraction to BEVDepth V2-99 | BEVDepth V2-99 · nuScenes val | mAP: 0.464 -> 0.459 | No clear effect |
| use max pooling instead of CollapseConv for feature reduction (ablation) | MatrixVT Prime Feature Extractor · nuScenes val Res-50 | NDS: 0.404 -> 0.415 | Improved |
| add refine sub-net after reduction (ablation) | MatrixVT Prime Feature Extractor · nuScenes val Res-50 | mAP: 0.328 -> 0.332 | Improved |
| add front-view position embedding (ablation) | MatrixVT Prime Feature Extractor · nuScenes val Res-50 | mAP: 0.332 -> 0.336 | Improved |
| use MatrixVT instead of Lift-Splat view transformers on CPU | view transformation · latency S1-S6 | speed: 4 to 8 times faster than LS-BEVDet | Improved |
| use MatrixVT instead of LS-BEVDepth for memory | view transformation · intermediate variables S1-S6 | memory footprint: 2 to 15 times less than LS-BEVDepth | Improved |
| use Ring and Ray decomposition with reformulated pipeline | MatrixVT · common setting C=80, Nd=112, WI=44 | calculation and memory: 46 times less calculation, 96% less memory | Improved |
| replace BEVDepth with MatrixVT on lane segmentation | BEV map segmentation · nuScenes val | IoU-Lane: 0.464 -> 0.448 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- BEVFormer — BEVFormer: Learning Bird’s-Eye-View Representation from Multi-Camera Images via Spatiotemporal Transformers (ECCV 2022) · better on 2 of 2 shared results
- PETR — PETR: Position Embedding Transformation for Multi-View 3D Object Detection (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
- EVT: Efficient View Transformation for Multi-Modal 3D Object Detection (ICCV 2025 · similarity 0.90)
- GeoBEV: Learning Geometric BEV Representation for Multi-view 3D Object Detection (AAAI 2025 · similarity 0.88)
- 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)
- Instance-Aware Multi-Camera 3D Object Detection with Structural Priors Mining and Self-Boosting Learning (AAAI 2024 · similarity 0.87)
- M-BEV: Masked BEV Perception for Robust Autonomous Driving (AAAI 2024 · 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.
