Research map / Detection and segmentation / 3D detection and driving
FSD-BEV: Foreground Self-Distillation for Multi-view 3D Object Detection
Zheng Jiang, Jinqing Zhang, Yanan Zhang, Qingjie Liu, Zhenghui HU, Baohui Wang, Yunhong Wang
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 Foreground Self-Distillation to BEVDepth (ablation) | BEVDepth ResNet50 · nuScenes val | mAP: 0.363 -> 0.394 | Improved |
| add Point Cloud Intensification on top of FSD (ablation) | FSD-BEV · nuScenes val | mAP: 0.394 -> 0.400 (NDS 0.516 -> 0.516) | Improved |
| add Multi-Scale Foreground Enhancement on top of FSD and PCI (ablation) | FSD-BEV · nuScenes val | NDS: 0.516 -> 0.526 | Improved |
| apply foreground segmentation when distilling (ablation) | FSD-BEV student · nuScenes val | mAP: 0.372 -> 0.393 | Improved |
| add frame combination to teacher point cloud (ablation) | FSD-BEV teacher · nuScenes val | mAP: 0.584 -> 0.605 | Improved |
| add pseudo point assignment after frame combination (ablation) | FSD-BEV teacher · nuScenes val | NDS: 0.650 -> 0.652 | No clear effect |
| raise MSFE foreground threshold from 0.10 to 0.30 (ablation) | MSFE · nuScenes val | mAP: 0.352 -> 0.344 | Hurt |
| use both F4 and F8 scales instead of a single scale (ablation) | MSFE · nuScenes val | mAP: 0.347 / 0.350 -> 0.352 | Improved |
| use self-distillation instead of cross-modal distillation with a larger backbone (ResNet101) | BEVDepth student · nuScenes val | mAP gain: DistillBEV +4.7% -> FSD-BEV +8.5% (ResNet50 to ResNet101) | Improved |
| use 2-frame FSD-BEV versus 8-frame StreamPETR on val | FSD-BEV ResNet101 · nuScenes val | mAP: 0.488 vs StreamPETR 0.504 mAP; NDS 0.589 vs 0.592 | Hurt |
Later papers that compared with this method
1 later accepted paper reports FSD-BEV as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Adaptive-Smooth LiDAR-Camera Knowledge Distillation with Heterogeneous Fusion for Multi-View 3D Object Detection (AAAI 2026 · better on 2 of 2 shared results)
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- SA-BEV — SA-BEV: Generating Semantic-Aware Bird's-Eye-View Feature for Multi-view 3D Object Detection (ICCV 2023) · better on 4 of 4 shared results
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
- DistillBEV: Boosting Multi-Camera 3D Object Detection with Cross-Modal Knowledge Distillation (ICCV 2023 · similarity 0.90)
- SimDistill: Simulated Multi-Modal Distillation for BEV 3D Object Detection (AAAI 2024 · similarity 0.90)
- BEV-MAE: Bird’s Eye View Masked Autoencoders for Point Cloud Pre-training in Autonomous Driving Scenarios (AAAI 2024 · similarity 0.88)
- FD3D: Exploiting Foreground Depth Map for Feature-Supervised Monocular 3D Object Detection (AAAI 2024 · similarity 0.87)
- Towards Generalizable Multi-Camera 3D Object Detection via Perspective Rendering (AAAI 2025 · similarity 0.87)
- CALICO: Self-Supervised Camera-LiDAR Contrastive Pre-training for BEV Perception (ICLR 2024 · similarity 0.87)
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.
