Research map / Point clouds / Segmentation and detection
Construct to Associate: Cooperative Context Learning for Domain Adaptive Point Cloud Segmentation
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 Cooperative Context Learning (CCL), a plug-in module for range-view point cloud domain adaptation that builds local kNN context embeddings with an MLP and noise masking, then associates them across domains via shared learnable prototypes with cross-attention.
Problem
Unsupervised domain adaptation for LiDAR point cloud semantic segmentation (synthetic-to-real), where geometric context (density, noise, beam angle) shifts across domains and existing methods ignore spatial context variation.
Approach
Combine local context embeddings (kNN neighbor weights from an MLP on relative coordinates, with noise-pixel masking) with global cross-domain context attention via shared learnable prototypes, where the global attention feeds back to refine local context modeling cooperatively.
Key results
- PLCA + CCL reaches 22.8% mIoU on SynLiDAR to SemKITTI (+4.2 over PLCA) and AdaptSeg + CCL reaches 23.3 / 18.1 on SynLiDAR to SemKITTI / nuScenes in ablation.
- SqzV1 + CCL achieves 28.2% mIoU on SynLiDAR to SemKITTI (+3.2 over SqzV1 at 25.0).
- CoSMix + CCL achieves 34.5 mIoU on SynLiDAR to SemanticKITTI with MinkowskiNet, versus 32.2 for CoSMix alone.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add CCL (context embedding + context attention) | SqueezeSegV1 / MMD / PLCA / AdaptSeg / CoSMix / LiDARNet backbones · SynLiDAR→SemKITTI and SynLiDAR→nuScenes | mIoU: +2.5 to +4.2 mIoU | Helped |
| add CCL | CoSMix (SalsaNext, nuScenes) · SynLiDAR→nuScenes | mIoU: +1.6 | Helped |
| remove both ConEmb and ConAtt (ablation) | AdaptSeg + CCL · SynLiDAR→SemKITTI | mIoU: 20.6 vs 23.3 full | Helped |
| remove both ConEmb and ConAtt (ablation) | AdaptSeg + CCL · SynLiDAR→nuScenes | mIoU: 13.0 vs 18.1 full | Helped |
| use domain-specific prototypes instead of domain-shared prototypes (ablation) | context attention · SynLiDAR→nuScenes | mIoU: 16.0 vs 18.1 | Hurt |
| insert module at 1st encoder block vs deeper positions (ablation) | CCL placement · SynLiDAR→nuScenes | mIoU: 18.1 vs 16.9 (3rd), 14.4 (5th), 14.3 (decoder) | Helped |
| add masking of noise samples in context embedding (ablation) | CCL context embedding · SynLiDAR→nuScenes | mIoU: +2.0 on nuScenes | Helped |
| increase neighbors k from 8 to 24 (grid 5x5) (ablation) | CCL context embedding · SynLiDAR→SemKITTI | mIoU: 23.3 to 27.2 (k=24), 29.5 (k=48), 27.8 (k=80); nuScenes 18.2, 19.9, 17.4 | Mixed |
| replace domain-shared with KPConv / Point Transformer / Meta-Kernel for context modeling (ablation) | context learning module comparison · SynLiDAR→SemKITTI | mIoU: 20.3 / 21.2 / 21.8 vs 23.3 | Hurt |
| add CCL (ablation) | SqueezeSegV3 and SalsaNext backbones · model capacity (FLOPS, params) | FLOPS and parameters: -0.1 to -0.5 FLOPS, -0.1 M params | No clear effect |
| add CCL without target-domain adaptation (source-only) | CCL alone · SynLiDAR→SemKITTI and nuScenes | mIoU: limited gains over source-only | No clear effect |
| apply 2D solutions (CBST, AdaptSeg, PLCA) to point clouds | 2D domain adaptation methods · SynLiDAR→SemKITTI | mIoU: attain negligible improvement over source-only | No clear effect |
Surprising findings
- Plugging CCL into 2D domain adaptation methods such as PLCA and AdaptSeg yields gains comparable to 3D methods, and 2D solutions alone attain negligible improvement over source-only on point clouds.
- Domain-specific prototypes perform worse than domain-shared prototypes, indicating cross-domain interaction through shared prototypes is needed for reducing the context gap.
Limitations
- CCL yields limited gains over source-only when no target data is used, since cross-domain association requires target scans.
- Performance is sensitive to the neighborhood size k; too large a receptive field harms results (k=48 best on SemKITTI but lower on nuScenes trade-offs).
- Evaluated only on range-view projections of LiDAR scans with a few synthetic-to-real transfers.
Experimental setup
Datasets: SynLiDAR SemanticKITTI nuScenes-lidarseg
Benchmarks: SynLiDAR to SemanticKITTI SynLiDAR to nuScenes
Metrics: mIoU
Baselines: CBST AdaptSeg CCM PLCA MMD SqueezeSegV2 ASM LiDARNet CoSMix SqueezeSegV1 SALUDA KPConv
Builds on: SqueezeSegV3 SalsaNext CoSMix PLCA AdaptSeg Lovasz-Softmax
Closest papers by meaning
- Complete & Label: A Domain Adaptation Approach to Semantic Segmentation of LiDAR Point Clouds (CVPR 2021 · similarity 0.90)
- BeyondMix: Leveraging Structural Priors and Long-Range Dependencies for Domain-Invariant LiDAR Segmentation (NeurIPS 2025 · similarity 0.90)
- Collaborative Learning for Semi-Supervised LiDAR Semantic Segmentation (ICML 2026 · similarity 0.88)
- Exploiting Local and Global Structure for Point Cloud Semantic Segmentation with Contextual Point Representations (NeurIPS 2019 · similarity 0.88)
- Exploring Scene Affinity for Semi-Supervised LiDAR Semantic Segmentation (CVPR 2025 · similarity 0.87)
- Mitigating Object Dependencies: Improving Point Cloud Self-Supervised Learning through Object Exchange (CVPR 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.