Research map / Point clouds / Point cloud classification
Mitigating Object Dependencies: Improving Point Cloud Self-Supervised Learning through Object Exchange
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes OESSL, a point cloud self-supervised learning framework that exchanges comparable-size objects between indoor scenes and learns context-aware object features for robustness to inter-object correlations.
Problem
Indoor point cloud models exploit strong human-habit-induced inter-object correlations instead of individual object patterns, and existing SSL methods preserve these dependencies, hurting robustness to context changes.
Approach
Exchange clusters of similar bounding-box size between scenes to break inter-object correlations without overlap, then train with object-pattern alignment across exchanged objects, contextual alignment for remaining clusters, and an auxiliary task predicting which points belong to exchanged objects.
Key results
- With 10% labels on ScanNet semantic segmentation, OESSL reaches 54.37 mIoU vs 48.99 from scratch and 53.85 for MSC.
- On S3DIS with 10% labels after ScanNet pre-training, OESSL achieves 49.22 mIoU vs 44.85 for MSC.
- On ScanNet-C robustness at replacement ratio 0.8, OESSL retains 54.43% of performance vs 52.92% for MSC.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use OESSL pre-training instead of MSC | MinkUnet · ScanNet semantic segmentation, 20% labels | mIoU: 60.47 -> 61.27 | Helped |
| Use OESSL pre-training instead of the strongest reported baseline | MinkUnet · S3DIS semantic segmentation, 10% labels | mIoU: 40.48 -> 49.22 | Helped |
| Use OESSL pre-training instead of MSC | MinkUnet · Synthia4D test semantic segmentation, 1% labels | mIoU: 66.42 -> 68.75 | Helped |
| Use OESSL pre-training instead of MSC | MinkUnet · Synthia4D validation semantic segmentation, 1% labels | mIoU: 71.58 -> 74.17 | Helped |
| Add object-pattern learning to context-cues learning (ablation) | OESSL · ScanNet semantic segmentation, 10% labels | mIoU: 53.12 -> 53.90 | Helped |
| Add the auxiliary task to context-cues and object-pattern learning (ablation) | OESSL · ScanNet semantic segmentation, 10% labels | mIoU: 53.90 -> 54.37 | Helped |
| Replace object exchange with Mix3D scene mixing (ablation) | OESSL pre-training · ScanNet semantic segmentation, 10% labels | mIoU: 54.37 -> 52.62 | Hurt |
| Use OESSL pre-training rather than the baseline pre-training configuration (ablation) | SPVCNN · ScanNet semantic segmentation, 10% labels | mIoU: 47.38 -> 49.02 | Helped |
Surprising findings
- Mix3D-style scene mixing as the augmentation for SSL underperforms MSC (52.62 vs 53.85 mIoU), since object overlaps disrupt feature structure.
Limitations
- Exchange proportion beta must be tuned; higher values cause object overlap and degrade performance.
- Gains over MSC are small at some label regimes (e.g., 100% ScanNet, 71.28 vs 71.00).
Experimental setup
Datasets: ScanNet S3DIS Synthia4D
Benchmarks: ScanNet S3DIS Synthia4D ScanNet-C
Metrics: mIoU Acc mAP AP@50 AP@25
Baselines: MSC DepthContrast STRL Mix3D training from scratch
Builds on: BYOL MinkUnet GraphCut Mix3D MSC
Closest papers by meaning
- Spatiotemporal Self-Supervised Learning for Point Clouds in the Wild (CVPR 2023 · similarity 0.89)
- Sonata: Self-Supervised Learning of Reliable Point Representations (CVPR 2025 · similarity 0.87)
- Construct to Associate: Cooperative Context Learning for Domain Adaptive Point Cloud Segmentation (CVPR 2024 · similarity 0.87)
- Relation3D : Enhancing Relation Modeling for Point Cloud Instance Segmentation (CVPR 2025 · similarity 0.87)
- PSA-SSL: Pose and Size-aware Self-Supervised Learning on LiDAR Point Clouds (CVPR 2025 · similarity 0.87)
- FreePoint: Unsupervised Point Cloud Instance Segmentation (CVPR 2024 · 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.