Text‐Inspired Feature Refinement for Few‐Shot 3D Point Cloud Semantic Segmentation

Abstract Few‐shot 3D point cloud semantic segmentation aims to segment novel categories in 3D scenes with scarce labeled support samples. Recent studies have predominantly followed the prototype‐based paradigm, focusing mainly on prototype enhancement. However, they typically neglect the significance of feature representation, which is pivotal to segmentation accuracy. In this work, we propose a framework to explicitly refine 3D features by exploiting geometric dependencies and leveraging semantic knowledge from text. Specifically, we first introduce a Channel‐Aware Feature Purification (CAFP) module, which mines latent global intra‐channel correlations to purify 3D features for stronger discriminability. Secondly, we propose a Text‐Anchored Modality Adaptation (TAMA) module to enhance inter‐class separability and intra‐class compactness of 3D features while alleviating the support‐query gap under textual guidance. Within this module, class‐specific text embeddings serve as semantic anchors to drive the adaptive alignment of features. Moreover, we design a Text‐Geometry (TG) loss to further transfer inter‐class semantic relationships from text to 3D geometry, facilitating more precise separations between foreground and background points. Extensive experiments demonstrate the superiority of our method, outperforming state‐of‐the‐art approaches by margins of 5.69% and 9.66% in mIoU on S3DIS and ScanNet, respectively. Our code is available at: https://github.com/TZH716/TIFR .

Authors

Institutions

Publication Details

Journal
Computer Graphics Forum
Published
2026-11-13
DOI
https://doi.org/10.1111/cgf.70635
Primary Topic
3D Shape Modeling and Analysis
Type
article
Field-Weighted Citation Impact
0.00

Funders

Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

Text‐Inspired Feature Refinement for Few‐Shot 3D Point Cloud Semantic Segmentation

Yong Zhao, Fan Zhong, Zihan Tang
2 links
Computer Graphics Forum
3D Shape Modeling and Analysis
article

Text‐Inspired Feature Refinement for Few‐Shot 3D Point Cloud Semantic Segmentation

Yong Zhao, Fan Zhong, Zihan Tang
article en

Abstract

Abstract Few‐shot 3D point cloud semantic segmentation aims to segment novel categories in 3D scenes with scarce labeled support samples. Recent studies have predominantly followed the prototype‐based paradigm, focusing mainly on prototype enhancement. However, they typically neglect the significance of feature representation, which is pivotal to segmentation accuracy. In this work, we propose a framework to explicitly refine 3D features by exploiting geometric dependencies and leveraging semantic knowledge from text. Specifically, we first introduce a Channel‐Aware Feature Purification (CAFP) module, which mines latent global intra‐channel correlations to purify 3D features for stronger discriminability. Secondly, we propose a Text‐Anchored Modality Adaptation (TAMA) module to enhance inter‐class separability and intra‐class compactness of 3D features while alleviating the support‐query gap under textual guidance. Within this module, class‐specific text embeddings serve as semantic anchors to drive the adaptive alignment of features. Moreover, we design a Text‐Geometry (TG) loss to further transfer inter‐class semantic relationships from text to 3D geometry, facilitating more precise separations between foreground and background points. Extensive experiments demonstrate the superiority of our method, outperforming state‐of‐the‐art approaches by margins of 5.69% and 9.66% in mIoU on S3DIS and ScanNet, respectively. Our code is available at: https://github.com/TZH716/TIFR .

Computer Graphics Forum
Shandong University (CN), Ocean University of China (CN)
National Natural Science Foundation of China, Zhejiang University, Natural Science Foundation of Shandong Province, State Key Laboratory of Computer Aided Design and Computer Graphics
3D Shape Modeling and Analysis
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon