COSTA leverages the domain gap through proven test-time adaptation, and groups each batch of target-domain points into a small set of semantic clusters based on the similarity distribution in the adapted feature space, and propagates high-confidence pseudo labels obtained from an open-vocabulary vision-language model to all points through cluster-level voting.
Abstract
Semantic segmentation of aerial point cloud is trapped in a generalization crisis under distinct domain shifts. While test-time adaptation offers a privacy-preserving and computationally efficient way to adapt pre-trained models to unlabeled target-domain data during inference, existing methods, bound to closed-set label assumptions and non-scalable point-wise segmentation pipelines, still struggle with semantic shifts. We ask: can we adapt any given pre-trained aerial point cloud segmentation model to a shifted target domain at the inference phase alone, without additional training, while segmenting target-specific categories beyond the source label space on demand? This paper introduces COSTA, which breaks this limitation by shifting from closed-set point-wise adaptation to cluster-centric open-set semantic propagation. Our core discovery is that, once effectively adapted at test time, the rich feature distribution of aerial point clouds can be distilled into a compact set of well-separated semantic centroids that are transferable across label spaces. COSTA leverages this to reformulate open-set semantic segmentation as a cluster-level propagating process: it first bridges the domain gap through proven test-time adaptation, then groups each batch of target-domain points into a small set of semantic clusters based on the similarity distribution in the adapted feature space, and finally propagates high-confidence pseudo labels obtained from an open-vocabulary vision-language model to all points through cluster-level voting. This cluster-centric paradigm enables test-time adaptation of aerial point clouds under significant domain gaps with mixed semantic shifts. With DALES as the source domain, COSTA enables on-demand segmentation across three aerial point cloud benchmarks with distinct domains and heterogeneous category spaces, achieving up to 70.09% mIoU under this new setting.
Class-incremental three-dimensional point cloud semantic segmentation requires models to learn newly introduced categories while preserving previously acquired knowledge without storing historical point clouds. This setting is challenged by representation drift during incremental optimization and semantic background shift caused by incomplete annotations of previously learned categories. To address these problems, this study proposes confidence-gated relational distillation, an exemplar-free teacher–student framework that combines feature-level relation preservation with semantic-level background correction. The relational component transfers normalized neighborhood-affinity distributions and weights each point according to teacher reliability, thereby reducing the influence of uncertain predictions. The background-compensation component reconstructs reliable old-class targets using class-specific thresholds and calibrates competition between previously learned and newly introduced classes. Experiments on the Stanford Large-Scale Three-Dimensional Indoor Spaces dataset and ScanNet show competitive performance across multiple incremental settings, with more consistent improvements on ScanNet. Under the ScanNet 10-1 protocol, the proposed method achieves an average mean intersection over union of 44.6% across eleven learning states and 33.3% at the final state. Under the same PointNet++ configuration, it also reduces training time and peak graphics processing unit memory. These results indicate that reliable relational transfer and adaptive background correction provide an effective balance between old-class retention and novel-class acquisition without introducing replay data or separate architectural branches during incremental optimization.
Lei Wang, Rongxiang Liu· Applied Sciences· 0 citations
Abstract. While deep learning has advanced ALS point cloud semantic segmentation and achieved impressive results, most methods rely on predefined label sets and lack ability to recognize arbitrary categories. Recently, the visual foundation models (VFMs) has garnered significant attention, due to remarkable zero-shot generalization capabilities by leveraging open-set knowledge. However, adapting these models to large-scale ALS point clouds remains largely unexplored and highly challenging. In addition, the frequent absence of well-aligned synchronously acquired images further hinders the application of 2D VFMs in ALS point clouds. To bridge these gaps, we developed a zero-shot, open-vocabulary semantic segmentation framework for ALS point clouds based on 2D-3D transfer, utilizing three types of VFMs. We employed a combination of VFMs, including source models pre-trained on natural imagery and models fine-tuned on remote sensing data, to investigate the generalization capabilities of VFMs in inherent domain gap between natural and aerial imagery. Besides, we further introduce an adaptive global view projection module that derives optimal virtual camera poses and field-of-view (FOV) from scene extents, effectively enabling the application of 2D VFMs even in the absence of original imagery. Quantitative evaluations on the Vaihingen dataset indicate that methods trained solely on natural images achieve segmentation accuracy scores of 72% (roof) and 59% (tree) for common classes but struggle with rare categories such as powerline. GSNET improves performance across most categories, highlighting importance of domain adaptation. Evaluation on the SUM dataset reveals that our approach effectively identifies large-scale urban elements (exceeding 60% precision for buildings) without high-quality, well-aligned imagery.
Yanghong Lin, Tianyu Li, Shudong Zhou et al.· The International Archives o...· 1 citation
This paper proposes a domain adaptation method for different lidars in point cloud semantic segmentation using semi-supervised learning and map annotation. This approach addresses the significant costs associated with acquiring point-level annotations in target domains. We first perform representation learning using large-scale unlabeled data. Subsequently, we conduct classification learning using a point cloud map integrated via coordinate transformations, eliminating the need to annotate individual frames. Experiments on our original TC dataset demonstrate that our method effectively bridges the domain gap from the SemanticKITTI dataset, improving the mean Intersection over Union (mIoU) by 48.6 points over a zero-shot baseline. Furthermore, our approach outperforms a strong baseline trained from scratch on the target domain by 7.2 points in mIoU. These results establish a cost-efficient solution for deploying diverse sensors.
So Nakanishi, Yoshitaka Hara, Y. Kuroda· 2026 IEEE/ASME International...· 0 citations
Semantic segmentation of forest point clouds is pivotal for automated forest inventory but is severely constrained by the scarcity of high-quality real-world annotations. Although synthetic forest point clouds provide a scalable alternative for training, models trained on synthetic data often suffer from pronounced performance degradation when transferred to real forest scenes, due to domain shifts coupled with extreme structural heterogeneity and long-tailed semantic distributions. To address these challenges, we propose ForestPlotUDA, an unsupervised domain adaptation (UDA) framework specifically designed for plot-level forest UAV light detection and ranging (LiDAR) semantic segmentation. Unlike generic adaptation methods that rely on global feature alignment or dense pseudo-labeling, ForestPlotUDA explicitly targets the severe class imbalance and plot-wise structural variability inherent in forest environments, where ecologically critical woody components are sparsely distributed and easily overwhelmed by dominant foliage points. The proposed framework integrates decoupled feature normalization and class-guided self-training, enabling stable cross-domain adaptation under highly imbalanced and sparse supervision. Experimental results on the FOR-instance benchmark demonstrate that ForestPlotUDA improves the mean intersection-over-union (mIoU) from 54.71% to 61.50%. Notably, using only five pseudo-labeled points per class, the intersection-over-union (IoU) of the challenging wood class increases from 22.82% to 40.90%, highlighting the effectiveness of the proposed approach for annotation-free forest UAV LiDAR analysis. Additional experiments on SegmentedForests further demonstrate that ForestPlotUDA consistently improves over direct synthetic-to-real (Syn2real) transfer on ground-based terrestrial laser scanning (TLS)/mobile laser scanning (MLS) forest point clouds. These results indicate that explicitly accounting for structural heterogeneity and class imbalance is critical for Syn2Real adaptation in forest point clouds, paving the way for fully automated and low-cost forest inventory systems across diverse forest ecosystems.
Jing Liu, Jing Huang, Di Wang· IEEE Transactions on Geoscie...· 0 citations
What if pathology foundation models could do more with less? GigaPath-Flash and GigaTIME-Flash cut computational demands while maintaining strong performance, opening the door to larger studies and broader exploration. The post GigaPath-Flash and GigaTIME-Flash: Toward population-scale discovery with efficient pathology foundation models appeared first on Microsoft Research.
MIT News · Artificial Intelligence· news.mit.eduAug 31, 2026
With millions of users across the world, Julia has been used to conduct cutting-edge research and to design new drugs, jet engines, heat pumps, and more.