Research map / Detection and segmentation / 3D detection and driving
MapDistill: Boosting Efficient Camera-based HD Map Construction via Camera-LiDAR Fusion Model Distillation
Xiaoshuai Hao, Ruikai Li, Hui Zhang, Rong Yin, Dingzhe Li, Sangil Jung, Seung-In Park, ByungIn Yoo, Haimei Zhao, Jing 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 all three distillation losses (relation, feature, head) to camera-only student (ablation) | MapTR ResNet18 student · nuScenes val | mAP: 45.9 -> 53.6 | Improved |
| add cross-modal relation distillation alone (ablation) | MapTR ResNet18 student · nuScenes val | mAP: 45.9 -> 48.8 | Improved |
| add map head distillation alone (ablation) | MapTR ResNet18 student · nuScenes val | mAP: 45.9 -> 49.0 | Improved |
| add dual-level feature distillation alone (ablation) | MapTR ResNet18 student · nuScenes val | mAP: 45.9 -> 48.4 | Improved |
| combine the three distillation losses (vs best pair) (ablation) | MapDistill · nuScenes val | mAP: 51.1 -> 53.6 | Improved |
| use cross-modal relation distillation instead of uni-modal (ablation) | MapDistill · nuScenes val | mAP: 52.0 -> 53.6 | Improved |
| add uni-modal relation to cross-modal relation (hybrid) (ablation) | MapDistill · nuScenes val | mAP: 53.6 -> 52.4 | Hurt |
| use dual-level instead of low-level only feature distillation (ablation) | MapDistill · nuScenes val | mAP: 52.7 -> 53.6 | Improved |
| use both classification and point losses in map head distillation (ablation) | MapDistill · nuScenes val | mAP: 51.9 -> 53.6 | Improved |
| use dual BEV transform with different 2D-to-BEV methods (GKT and LSS) (ablation) | student model · nuScenes val | mAP: 49.1 -> 53.6 | Improved |
| use same 2D-to-BEV method in both dual BEV branches (ablation) | student model · nuScenes val | mAP: 49.1 -> 50.9 | Mixed |
| use camera-LiDAR fusion teacher instead of camera-only or LiDAR-only teacher | MapDistill · nuScenes val | mAP: 48.0 / 50.1 -> 53.6 | Improved |
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.
- BEVDistill — BEVDistill: Cross-Modal BEV Distillation for Multi-View 3D Object Detection (ICLR 2023) · 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
- MapExpert: Online HD Map Construction with Simple and Efficient Sparse Map Element Expert (AAAI 2025 · similarity 0.87)
- The Road Ahead in Autonomous Driving: The KITScenes Multimodal Dataset (NeurIPS 2026 · similarity 0.86)
- Distilling Monocular Foundation Model for Fine-grained Depth Completion (CVPR 2025 · similarity 0.86)
- SimDistill: Simulated Multi-Modal Distillation for BEV 3D Object Detection (AAAI 2024 · similarity 0.85)
- CRKD: Enhanced Camera-Radar Object Detection with Cross-modality Knowledge Distillation (CVPR 2024 · similarity 0.85)
- SafeMap: Robust HD Map Construction from Incomplete Observations (ICML 2025 · 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.
