Query-aware Hub Prototype Learning for Few-Shot 3D Point Cloud Semantic Segmentation
YiLin Zhou, Lili Wei, Zheming Xu, Ziyi Chen, Congyan Lang
TL;DR
This paper introduces a query-aware hub prototype learning approach for few-shot 3D point cloud segmentation, explicitly modeling support-query relations to improve prototype relevance and segmentation accuracy.
Contribution
It proposes a novel QHP method with HPG and PDO modules to generate query-relevant prototypes and optimize their distribution, addressing prototype bias in few-shot segmentation.
Findings
Significant performance improvements on S3DIS and ScanNet datasets.
Effective reduction of prototype bias and improved generalization.
Enhanced segmentation accuracy in few-shot scenarios.
Abstract
Few-shot 3D point cloud semantic segmentation (FS-3DSeg) aims to segment novel classes with only a few labeled samples. However, existing metric-based prototype learning methods generate prototypes solely from the support set, without considering their relevance to query data. This often results in prototype bias, where prototypes overfit support-specific characteristics and fail to generalize to the query distribution, especially in the presence of distribution shifts, which leads to degraded segmentation performance. To address this issue, we propose a novel Query-aware Hub Prototype (QHP) learning method that explicitly models semantic correlations between support and query sets. Specifically, we propose a Hub Prototype Generation (HPG) module that constructs a bipartite graph connecting query and support points, identifies frequently linked support hubs, and generates query-relevant…
Peer Reviews
Decision·Submitted to ICLR 2026
1. It is interesting to introduce the concept of hubs into few-shot 3D point cloud semantic segmentation. By selecting the parts of the support point cloud closest to the query to generate prototypes, the resulting prototypes better align with the distribution of the query point cloud. 2. The experimental evaluation is sufficiently comprehensive and convincingly supports the effectiveness of the proposed QHP.
1. The experiment in Table 5 suggests that QHP is relatively sensitive to hyperparameter choices, raising concerns about its generalization across different datasets. This sensitivity may also explain why the performance gains on ScanNet are less pronounced than those on S3DIS. To address this issue, please verify whether substantially different hyperparameters are required for ScanNet. 2. The temperature parameter in Equation 8 has not been subjected to ablation analysis, despite being a criti
- The introduction of “hubness” for prototype generation in FS-3DSeg is novel and provides a new perspective on addressing support-query misalignment. - Extensive experiments: Evaluations on S3DIS and ScanNet across 1-shot and 5-shot settings are comprehensive. Ablation studies and parameter sensitivity analyses support the effectiveness of the proposed modules. - Clear writing and figures: The paper is well written, and figures (e.g., Figure 1, Figure 2, Figure 3) effectively illustrate the key
- Limited comparisons to recent or stronger baselines: The paper lacks comparisons with more recent state-of-the-art methods beyond COSeg and QGE/QGPA. Recent approaches that incorporate transformer-based meta learners, distillation, or prompt-based adaptation are not discussed or evaluated. - Unclear generalization and robustness of hub mining: The “hubness” concept is somewhat heuristic and heavily relies on k-NN and purity thresholds. There is limited discussion or analysis on whether hub min
The motivation for addressing prototype bias from using only support features is interesting. The experimental comparison is sufficient. Stepwise ablations (Baseline → +HPG → +PDO) indicate incremental improvements
1. The abstract and introduction state that “existing metric-based prototype learning methods generate prototypes solely from the support set, without considering their relevance to query data.” This is not generally accurate. A substantial body of transductive / query-guided work explicitly leverages unlabeled query features to refine or align prototypes with the query distribution, e.g., support–query alignment and query-aware refinement [1, 2, 3]. Because this statement underpins the paper’s
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
Topics3D Shape Modeling and Analysis · Robotics and Sensor-Based Localization · Advanced Neural Network Applications