Research map / Point clouds / Segmentation and detection
LiDAR Distillation: Bridging the Beam-Induced Domain Gap for 3D Object Detection
Yi Wei, Zibu Wei, Yongming Rao, Jiaxin Li, Jie Zhou, Jiwen Lu
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 point cloud density alignment (pseudo low-beam training data) (ablation) | PointPillars · KITTI 32-beam target (moderate) | AP_3D: 65.91 -> 67.01 | Improved |
| add knowledge distillation on top of density alignment (ablation) | PointPillars · KITTI 32-beam target (moderate) | AP_3D: 67.01 -> 70.15 | Improved |
| initialize student from pretrained teacher weights (ablation) | PointPillars · KITTI 32-beam target (moderate) | AP_3D: 69.01 -> 70.15 | Improved |
| use progressive 32-beam teacher instead of 64-beam teacher (ablation) | PointPillars · KITTI 16-beam target (moderate) | AP_3D: 58.80 -> 59.87 | Improved |
| use progressive 32-beam teacher instead of directly trained 32-beam teacher (ablation) | PointPillars · KITTI 16-beam target (moderate) | AP_3D: 58.74 -> 59.87 | Improved |
| mimic ROI regions instead of the whole BEV map (ablation) | PointPillars · KITTI 32-beam target (moderate) | AP_3D: 69.36 -> 70.15 | Improved |
| mimic ROI regions instead of groundtruth box regions only (ablation) | PointPillars · KITTI 32-beam target (moderate) | AP_3D: 69.99 -> 70.15 | Improved |
| use beam clustering downsampling instead of range-image interpolation upsampling (ablation) | PointPillars · KITTI 32-beam target (moderate) | AP_3D: 54.56 -> 70.15 | Improved |
| use downsampling instead of super-resolution upsampling (ablation) | PointPillars · KITTI 32-beam target (moderate) | AP_3D: 62.58 -> 70.15 | Improved |
| apply general UDA method ST3D | PointPillars · KITTI 32-beam target (moderate) | AP_3D: 65.91 -> 57.57 | Hurt |
| align source beams to 16* equivalent beams of target (ablation) | SECOND-IoU · Waymo -> nuScenes | AP_3D: 17.24 -> 22.86 | Improved |
| align source density to target (sacrifices source domain performance) (ablation) | SECOND-IoU · Waymo (source domain) | AP_3D: 54.01 -> 39.97 | Hurt |
Closest papers by meaning
- Density-Insensitive Unsupervised Domain Adaption on 3D Object Detection (CVPR 2023 · similarity 0.91)
- Weak-to-Strong 3D Object Detection with X-Ray Distillation (CVPR 2024 · similarity 0.90)
- SimDistill: Simulated Multi-Modal Distillation for BEV 3D Object Detection (AAAI 2024 · similarity 0.89)
- Three Pillars Improving Vision Foundation Model Distillation for Lidar (CVPR 2024 · similarity 0.89)
- GBlobs: Explicit Local Structure via Gaussian Blobs for Improved Cross-Domain LiDAR-based 3D Object Detection (CVPR 2025 · similarity 0.89)
- An Empirical Study of the Generalization Ability of Lidar 3D Object Detectors to Unseen Domains (CVPR 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.
