CVPR Findings2026
SCOPE: Spatially Ordered Continual Learning for 3D Segmentation
Wenhao Xu, Huaidong Zhang, Weipeng Zhang, Qianle Zhang, Shengfeng He
Abstract
Modern 3D semantic segmentation performs well in static settings, but adapting it to continual learning remains challenging. When new classes arrive without full data replay, models are prone to catastrophic forgetting and pseudo label drift. Unlike 2D images, point clouds are unordered and sparsely sampled, making neighborhood construction unstable and hindering the transfer of 2D continual-learning strategies. Existing 3D methods often rely on dynamic kNN graphs whose locality is sensitive to density, occlusion, and sampling variations, leading to inconsistent relational cues across incremental steps. To address these issues, we introduce SCOPE (Spatially Ordered Continual Point Embeddings), a framework that converts point clouds into topology-preserving spatial orders, enabling sequencebased processing while retaining geometric fidelity. On these ordered views, we propose Spatial-Order Relational Modeling (SORM), which aggregates context through rankbased linear neighborhoods that offer more stable locality than kNN. We further present Serialized Uncertainty Refinement (SUR), which combines intra-point entropy with inter-order consensus to form a traversal-consistent teacher for pseudo-label refinement. By making spatial order central, SCOPE provides backbone-agnostic relational modeling and uncertainty calibration. Extensive experiments on S3DIS and ScanNet, covering both sequence-oriented transformer backbone and classic graph-based architecture, show that SCOPE consistently improves the stabilityplasticity trade-off, reduces forgetting, and outperforms both adapted 2D methods and strong 3D baselines.
Ask Lune
Know this paper in minutes, not an afternoon.
Lune can read every page of it and the top-tier papers that cite it, and gives you its claims, its evidence and its limits, each one cited to the paper it comes from. No skimming, no AI slop.
Lineage
Builds on18
- Dark Experience for General Continual Learning: a Strong, Simple BaselinePietro Buzzega, Matteo Boschini, Angelo Porrello, Davide Abati et al.NeurIPS 2020 · 1,494 citations
- PseudoSeg: Designing Pseudo Labels for Semantic SegmentationYuliang Zou, Zizhao Zhang, Han Zhang, Chun-Liang Li et al.ICLR 2021 · 364 citations
- PointGPT: Auto-regressively Generative Pre-training from Point CloudsGuangyan Chen, Meiling Wang, Yi Yang, Kai Yu et al.NeurIPS 2023 · 219 citations
- SSUL: Semantic Segmentation with Unknown Label for Exemplar-based Class-Incremental LearningSungmin Cha, Beomyoung Kim, Youngjoon Yoo, Taesup MoonNeurIPS 2021 · 139 citations
- Mining Unseen Classes via Regional Objectness: A Simple Baseline for Incremental SegmentationZekang Zhang, Guangyu Gao, Zhiyuan Fang, Jianbo Jiao et al.NeurIPS 2022 · 53 citations
Related papers
- SCOPE: Scene-Contextualized Incremental Few-Shot 3D SegmentationVishal Thengane, Zhaochong An, Tianjin Huang, Son Lam Phung, Abdesselam Bouzerdoum, Lu Yin, Na Zhao, Xiatian ZhuCVPR Findings 2026
- Geometry and Uncertainty-Aware 3D Point Cloud Class-Incremental Semantic SegmentationYuwei Yang, Munawar Hayat, Zhao Jin, Chao Ren et al.CVPR 2023
- GeoGuide: Hierarchical Geometric Guidance for Open-Vocabulary 3D Semantic SegmentationXujing Tao, Chuxin Wang, Yubo Ai, Zhixin Cheng et al.CVPR 2026 · 3 citations
- Label-Guided Knowledge Distillation for Continual Semantic Segmentation on 2D Images and 3D Point CloudsZe Yang, Ruibo Li, Evan Ling, Chi Zhang et al.ICCV 2023 · 23 citations
- PointCSP: Cross-Sample Semantic Propagation and Stability Preservation in Self-Supervised Point Cloud LearningXinxing Yu, Ajian Liu, Sunyuan Qiang, Hui Ma et al.CVPR 2026 · 1 citation
