[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-81542-en":3,"doc-seo-81542-105":29,"detail-sidebar-cat-0-en-105":91},{"code":4,"msg":5,"data":6},0,"success",{"doc_id":7,"user_id":8,"nickname":9,"user_avatar":10,"doc_module":4,"category_id":11,"category_name":12,"doc_title":13,"doc_description":14,"doc_content":15,"file_id":16,"file_url":17,"file_type":18,"file_size":19,"view_count":20,"is_deleted":4,"is_public":20,"is_downloadable":20,"audit_status":20,"page_count":21,"language":22,"language_code":23,"site_id":24,"html_lang":23,"table_of_contents":25,"faqs":26,"seo_title":13,"seo_description":14,"update_tm":27,"read_time":28},81542,687197100911,"Himbo","https://ap-avatar.wpscdn.com/avatar/a000239b6f1da00475?x-image-process=image/resize,m_fixed,w_180,h_180&k=1782698725881665579",8,"Research & Report","White Aggregation and Restoration for Few-shot 3D Point Cloud Semantic Segmentation","Few-shot 3D point cloud semantic segmentation (FS-PCS) predicts per-point labels for unlabeled point clouds using only a few labeled examples. Existing approaches build class prototypes via farthest point sampling (FPS), but prototype generation becomes unstable because performance is highly sensitive to FPS-induced variations, while the prototype construction process remains insufficiently studied. The work proposes a deterministic attention-based method using a White Aggregation and Restoration Module (WARM), aligning token distributions via whitening and then restoring them via coloring. Experiments show state-of-the-art results on S3DIS and competitive performance on ScanNet.","arXiv :2509 . 13907v3 [ cs .CV] 10 Jul 2026  \nWhite Aggregation and Restoration for Few-shot 3D Point Cloud Semantic Segmentation  \nJiyun Im, SuBeen Lee, Miso Lee, and Jae-Pil Heo*  \nSungkyunkwan University  \n{bbangsil0110, leesb7426, dlalth557, [jaepilheo}@skku.edu](jaepilheo}@skku.edu)  \nAbstract. Few-shot 3D Point Cloud Semantic Segmentation (FS-PCS) aims to predict per-point labels for an unlabeled point cloud, given only a few labeled examples. To extract representations from the limited labeled set, existing methods have constructed prototypes with Farthest Point Sampling (FPS) . However, we found that this convention results in performance instability due to its sensitivity to FPS-induced variations, while the prototype generation process remains underexplored in the field. This motivates us to investigate deterministic prototype generation method based on attention mechanism. Despite its potential, we found that vanilla attention module suffers from the distributional gap between prototypical tokens and support features. To overcome this, we provide a simple approach, White Aggregation and Restoration Module (WARM), which resolves the misalignment by wrapping cross-attention with whitening and coloring transformations. Specifically, whitening aligns the features to tokens before the attention process, and coloring subsequently restores the original distribution to the attended tokens. This design enables robust attention, thereby generating prototypes that capture the semantic relationships in support features. WARM achieves state-ofthe-art performance with a significant margin on the S3DIS dataset, and competitive performance on the ScanNet dataset. Further experiments demonstrate its effectiveness in deterministic prototype generation. Code is publicly available at: [https://github.com/JiyunIm00/WARM.git](https://github.com/JiyunIm00/WARM.git)  \nKeywords: Few-shot Learning · 3D Semantic Segmentation · Attention  \n1 Introduction  \nUnderstanding the semantics of 3D point clouds has become crucial as its applications have advanced [7, 34] . Although recent methods have achieved remarkable performance, they rely on large amounts of labeled data, which requires expensive labor [4, 10] . To alleviate this data reliance, Few-shot 3D Point Cloud Semantic Segmentation (FS-PCS) was introduced [39] . It aims to segment unseen novel classes in unlabeled point clouds, referred to as the query set, using a small number of labeled examples, referred to as the support set.  \n* Corresponding author  \n2 J. Im et al.  \nTable 1: mIoU (%) distribution resulting from random seeds. Performance is evaluated under the 1-way 1-shot setting on the first split of S3DIS [4] . ‘FPS + min-dist.’ denotes a simple baseline that assigns labels based on the minimum distance to FPS-constructed class prototypes. The distribution of mIoUs show the sensitivity of FS-PCS models to FPS results, that even the complex segmentation structure [3] falls short to generalize. On the other hand, attention-based WARM shows much stability.  \n\n| Method | Max | Min | Mean | Std. |\n| --- | --- | --- | --- | --- |\n| COSeg [3] | 52.86 | 37.99 | 45.67 | 2.41 |\n| FPS + min-dist. | 52.14 | 46.86 | 49.48 | 0.86 |\n| WARM | 60.16 | 60.16 | 60.16 | - |\n\n\n|  Prototypical Tokens  Point Cloud Features  Centroid of  |  |  |\n| --- | --- | --- |\n| \u003Cbr>\u003Cbr>Naïve Cross-Attn |  | \u003Cbr>\u003Cbr>WARM (Ours) |\n|  |  |  |\n\n(a) Visualization of Feature Space  \n1.0  \n0.5  \n0.0  \n (b) Sharpness of Attention Weight  Tokens  \nFig. 1: Comparison of naïve cross-attention and WARM in prototype construction. (a) t-SNE visualization of the alignment between prototypical tokens (queries) and point cloud features (keys) . (b) Attention sharpness measured via attention entropy [37], averaged for each token.  \nTo fully utilize the information from the support set, previous studies have constructed class prototypes [2,3,14,26,39] . It is widely adopted in other few-shot downstream tasks, where numerous studies f","cbCaivHMcbVqkjXF","https://ap.wps.com/l/cbCaivHMcbVqkjXF","pdf",1428251,1,22,"English","en",105,"# Introduction\n## Motivation for deterministic prototype generation\n## Limitations of FPS-based prototypes\n## Limitations of vanilla cross-attention\n# White Aggregation and Restoration Module (WARM)\n## Whitening and coloring for distribution alignment\n## Robust prototype construction\n# Experiments and results","[{\"question\":\"What problem does FS-PCS address?\",\"answer\":\"FS-PCS targets per-point semantic labeling for an unlabeled 3D point cloud using only a few labeled support examples to infer novel classes in the query set.\"},{\"question\":\"Why can FPS-based prototype generation hurt performance stability?\",\"answer\":\"FPS-derived prototypes depend on the initially selected points, causing inconsistent prototypes and fluctuations in segmentation performance, which can be amplified by later sophisticated modules.\"},{\"question\":\"How does WARM improve attention-based deterministic prototype generation?\",\"answer\":\"WARM wraps cross-attention with whitening and coloring transformations: whitening aligns prototype tokens with support features before attention, and coloring restores the attended tokens’ distribution to better capture semantic relationships.\"}]",1784174180,55,{"code":4,"msg":30,"data":31},"ok",{"site_id":24,"language":23,"slug":32,"title":13,"keywords":33,"description":14,"schema_data":34,"social_meta":86,"head_meta":88,"extra_data":90,"updated_unix":27},"white-aggregation-and-restoration-for-few-shot-3d-point-cloud-semantic-segmentation","",{"@graph":35,"@context":85},[36,53,68],{"@type":37,"itemListElement":38},"BreadcrumbList",[39,43,47,50],{"item":40,"name":41,"@type":42,"position":20},"https://docshare.wps.com","Home","ListItem",{"item":44,"name":45,"@type":42,"position":46},"https://docshare.wps.com/document/","Document",2,{"item":48,"name":12,"@type":42,"position":49},"https://docshare.wps.com/document/research-report/",3,{"item":51,"name":13,"@type":42,"position":52},"https://docshare.wps.com/document/white-aggregation-and-restoration-for-few-shot-3d-point-cloud-semantic-segmentation/81542/",4,{"url":51,"name":13,"@type":54,"author":55,"headline":13,"publisher":57,"fileFormat":60,"inLanguage":23,"description":14,"dateModified":61,"datePublished":62,"encodingFormat":60,"isAccessibleForFree":63,"interactionStatistic":64},"DigitalDocument",{"name":9,"@type":56},"Person",{"url":40,"name":58,"@type":59},"DocShare","Organization","application/pdf","2026-07-25","2026-07-16",true,{"@type":65,"interactionType":66,"userInteractionCount":20},"InteractionCounter",{"@type":67},"ViewAction",{"@type":69,"mainEntity":70},"FAQPage",[71,77,81],{"name":72,"@type":73,"acceptedAnswer":74},"What problem does FS-PCS address?","Question",{"text":75,"@type":76},"FS-PCS targets per-point semantic labeling for an unlabeled 3D point cloud using only a few labeled support examples to infer novel classes in the query set.","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"Why can FPS-based prototype generation hurt performance stability?",{"text":80,"@type":76},"FPS-derived prototypes depend on the initially selected points, causing inconsistent prototypes and fluctuations in segmentation performance, which can be amplified by later sophisticated modules.",{"name":82,"@type":73,"acceptedAnswer":83},"How does WARM improve attention-based deterministic prototype generation?",{"text":84,"@type":76},"WARM wraps cross-attention with whitening and coloring transformations: whitening aligns prototype tokens with support features before attention, and coloring restores the attended tokens’ distribution to better capture semantic relationships.","https://schema.org",{"og:url":51,"og:type":87,"og:title":13,"og:site_name":58,"og:description":14},"article",{"robots":89,"canonical":51},"index,follow",{"doc_id":7,"site_id":24},{"code":4,"msg":5,"data":92},[93,97,101,105,110,115,120,123,128,131,135],{"id":20,"doc_module":4,"doc_module_name":45,"category_name":94,"show_sort_weight":95,"slug":96},"Story & Novel",90,"story-novel",{"id":46,"doc_module":4,"doc_module_name":45,"category_name":98,"show_sort_weight":99,"slug":100},"Literature",80,"literature",{"id":52,"doc_module":4,"doc_module_name":45,"category_name":102,"show_sort_weight":103,"slug":104},"Exam",70,"exam",{"id":106,"doc_module":4,"doc_module_name":45,"category_name":107,"show_sort_weight":108,"slug":109},5,"Comic",60,"comic",{"id":111,"doc_module":4,"doc_module_name":45,"category_name":112,"show_sort_weight":113,"slug":114},6,"Technology",50,"technology",{"id":116,"doc_module":4,"doc_module_name":45,"category_name":117,"show_sort_weight":118,"slug":119},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":45,"category_name":12,"show_sort_weight":121,"slug":122},30,"research-report",{"id":124,"doc_module":4,"doc_module_name":45,"category_name":125,"show_sort_weight":126,"slug":127},9,"Religion & Spirituality",20,"religion-spirituality",{"id":126,"doc_module":4,"doc_module_name":45,"category_name":129,"show_sort_weight":126,"slug":130},"World Cup","world-cup",{"id":132,"doc_module":4,"doc_module_name":45,"category_name":133,"show_sort_weight":132,"slug":134},10,"Lifestyle","lifestyle",{"id":136,"doc_module":4,"doc_module_name":45,"category_name":137,"show_sort_weight":106,"slug":138},19,"General","general"]