Research map / Point clouds / Segmentation and detection
Self-Supervised Learning of LiDAR 3D Point Clouds via 2D-3D Neural Calibration
Yifan Zhang, Junhui Hou, Siyu Ren, Jinjian Wu, Yixuan Yuan, Guangming Shi
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 learnable transformation alignment to the baseline pretext task (ablation) | NCLR pre-training · nuScenes semantic seg (1% labels, ablation split) | mIoU: 35.90 -> 36.52 | Improved |
| add overlapping area detection on top of learnable transformation alignment (ablation) | NCLR pre-training · nuScenes semantic seg (1% labels, ablation split) | mIoU: 36.52 -> 37.13 | Improved |
| add end-to-end pose estimation (soft matching, differentiable EPnP, pose loss) (ablation) | NCLR pre-training · nuScenes semantic seg (1% labels, ablation split) | mIoU: 37.13 -> 37.92 | Improved |
| use learnable transformation alignment instead of plain cosine similarity (ablation) | image-point feature similarity · SemanticKITTI val | feature similarity: 0.55 ± 0.17 -> 0.69 ± 0.13 | Improved |
| replace hard matching with soft matching (cosine similarity) (ablation) | point-to-pixel matching · SemanticKITTI feature matching | matching accuracy (<5 px): 69.22 -> 74.27 | Improved |
| add learnable transformation alignment (hard matching) (ablation) | point-to-pixel matching · SemanticKITTI feature matching | matching accuracy (<5 px): 69.22 -> 83.18 | Improved |
| combine learnable alignment with soft matching (ablation) | point-to-pixel matching · SemanticKITTI feature matching | matching accuracy (<5 px): 83.18 -> 90.28 | Improved |
| use NCLR pre-training instead of no pre-training with 0.1% labels | MinkUNet · nuScenes semantic seg | mIoU: 21.6 -> 26.6 | Improved |
| use NCLR pre-training instead of training from scratch with 0.1% labels | MinkUNet · SemanticKITTI semantic seg | mIoU: 30.0 -> 39.2 | Improved |
| NCLR pre-training with full labels versus no pre-training | MinkUNet · SemanticKITTI semantic seg (100% labels) | mIoU: 62.7 -> 63.9 | Improved |
| PointContrast/DepthContrast pre-training with full labels | MinkUNet · nuScenes semantic seg (100% labels) | mIoU: 71.2 -> 71.1 | No clear effect |
| NCLR pre-training for detection with 10% KITTI labels | PV-RCNN · KITTI3D (10% labels) | mAP: 64.04 -> 66.93 | 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.
- ProposalContrast — ProposalContrast: Unsupervised Pre-training for LiDAR-Based 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
- Image-to-Lidar Self-Supervised Distillation for Autonomous Driving Data (CVPR 2022 · similarity 0.91)
- PSA-SSL: Pose and Size-aware Self-Supervised Learning on LiDAR Point Clouds (CVPR 2025 · similarity 0.91)
- Temporal Consistent 3D LiDAR Representation Learning for Semantic Perception in Autonomous Driving (CVPR 2023 · similarity 0.90)
- GhostPoint: Self-Supervised Representation Learning by Hallucinating Occluded LiDAR Structure (ECCV 2026 · similarity 0.90)
- GraphAlign: Enhancing Accurate Feature Alignment by Graph matching for Multi-Modal 3D Object Detection (ICCV 2023 · similarity 0.89)
- VLM2Scene: Self-Supervised Image-Text-LiDAR Learning with Foundation Models for Autonomous Driving Scene Understanding (AAAI 2024 · similarity 0.89)
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.
