Research map / Point clouds / Segmentation and detection
Novel Class Discovery for Point Cloud Segmentation via Joint Learning of Causal Representation and Reasoning
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 a causal novel-class-discovery method for 3D point cloud segmentation that removes confounders via adversarial causal prototypes and reasons over a base-to-novel causal graph with GCN pseudo-labeling.
Problem
Novel class discovery for point cloud segmentation (3D-NCD): segment unlabeled novel classes using only labeled base-class supervision, where shortcut features and statistical correlations mislead base representations and novel-class inference.
Approach
Model the task with a structural causal model; learn confounder-free causal representation prototypes for base classes via adversarial training (ICM principle), then build a graph with base causal prototypes and novel prototypes whose edges are pruned, direction-constrained, and propagated with a GCN to produce novel pseudo-labels.
Key results
- On SemanticPOSS split 3, the method outperforms DASL on novel classes by 10.2%, with cone-stone IoU of 36.2% where DASL and NOPS get 0.
- On SemanticKITTI, the method reaches the highest overall mIoU in all four splits; e.g., split 1 novel-class improvement of 4.9% over DASL and car IoU of 82.6%.
- On PASCAL-5i the method averages 61.2 vs. EUMS 59.1; on COCO-20i it averages 27.00 vs. EUMS 26.81.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the full causal representation and reasoning method rather than DASL | Novel class discovery · SemanticPOSS split 3 | Novel mIoU: 17.7 -> 27.9 | Helped |
| Use the full causal representation and reasoning method rather than DASL | Novel class discovery · SemanticKITTI split 1 | Novel mIoU: 28.7 -> 33.6 | Helped |
| Use the full causal representation and reasoning method rather than SNOPS | Novel class discovery · SemanticKITTI split 2 | Novel mIoU: 20.1 -> 18.5 | Hurt |
| Use the full causal representation and reasoning method rather than EUMS | 2D novel class discovery · PASCAL-5^i Avg | mIoU: 59.1 -> 61.2 | Helped |
| Use the full causal representation and reasoning method rather than EUMS | 2D novel class discovery · COCO-20^i Avg | mIoU: 26.81 -> 27.00 | Helped |
| Add CRP (ablation) | Baseline MinkowskiUNet-34C · SemanticPOSS split 0 | Novel mIoU: 38.3 -> 41.6 | Helped |
| Add CRP (ablation) | Baseline MinkowskiUNet-34C · SemanticPOSS overall | Novel mIoU: 24.7 -> 25.9 | Helped |
| Add CRG using traditional clustering prototypes as graph nodes (ablation) | Baseline MinkowskiUNet-34C · SemanticPOSS split 0 | Novel mIoU: 38.3 -> 45.5 | Helped |
| Add CRG using traditional clustering prototypes as graph nodes (ablation) | Baseline MinkowskiUNet-34C · SemanticPOSS overall | Novel mIoU: 24.7 -> 28.8 | Helped |
| Add CRP and CRG (ablation) | Baseline MinkowskiUNet-34C · SemanticPOSS split 0 | Novel mIoU: 38.3 -> 47.4 | Helped |
| Add CRP and CRG (ablation) | Baseline MinkowskiUNet-34C · SemanticPOSS overall | Novel mIoU: 24.7 -> 30.1 | Helped |
| Add GCPL using GCN-generated pseudo-labels (ablation) | Baseline + CRP + CRG · SemanticPOSS split 0 | Novel mIoU: 47.4 -> 51.2 | Helped |
Limitations
- Novel-class features are not used for causal representation learning because pseudo-labels may be biased or noisy.
- Confounders are only handled implicitly via adversarial training since they are unobservable.
Experimental setup
Datasets: SemanticKITTI SemanticPOSS PASCAL-5i COCO-20i
Benchmarks: SemanticKITTI SemanticPOSS PASCAL-5i COCO-20i
Metrics: IoU mIoU
Baselines: EUMS NOPS SNOPS DASL
Builds on: NOPS DASL Sinkhorn-Knopp MinkowskiUNet
Closest papers by meaning
- X-3D: Explicit 3D Structure Modeling for Point Cloud Recognition (CVPR 2024 · similarity 0.87)
- Interpretable3D: An Ad-Hoc Interpretable Classifier for 3D Point Clouds (AAAI 2024 · similarity 0.86)
- Hyperbolic Uncertainty-Aware Few-Shot Incremental Point Cloud Segmentation (CVPR 2025 · similarity 0.86)
- LogoSP: Local-global Grouping of Superpoints for Unsupervised Semantic Segmentation of 3D Point Clouds (CVPR 2025 · similarity 0.86)
- Weakly Supervised Point Cloud Semantic Segmentation via Artificial Oracle (CVPR 2024 · similarity 0.86)
- Rethinking Few-shot 3D Point Cloud Semantic Segmentation (CVPR 2024 · similarity 0.86)
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.