Skip to content
Book Open access

SegmentGDA: Training-Free Few-Shot Segmentation via Gaussian Discriminant Analysis on Vision Foundation Features

Nov 2026 · Proceedings of the 34th ACM International Conference on Multimedia · 0 citations · 2 references

Abstract

Few-shot segmentation aims to segment novel object categories given only a handful of annotated examples. Existing approaches either rely on meta-trained specialist models whose learned matching functions generalize poorly across domains or adopt training-free pipelines based on point-wise feature similarity that treat each patch embedding independently and fail to capture the global statistical structure of the target concept. In this paper, we present SegmentGDA, a fully training-free framework that recasts few-shot segmentation as probabilistic inference in the representation space of frozen vision foundation models. SegmentGDA models foreground and background support features as class-conditional Gaussian distributions and applies Gaussian Discriminant Analysis to classify each query patch via Bayesian decision rules, producing a dense posterior probability map that aggregates global class-level statistics across all support samples and suppresses noisy correspondences more effectively than deterministic local matching. The posterior map is further converted into spatially diverse positive and negative point prompts for the Segment Anything Model through an iterative refinement loop, yielding high-resolution segmentation masks without any task-specific training or fine-tuning. Extensive experiments across four standard benchmarks demonstrate the effectiveness of the proposed framework, underscoring its potential for real-world applications. These results highlight distribution-based modeling as a promising approach for adapting vision foundation models to few-shot dense prediction. Code is available at https://github.com/wavachao/SegmentGDA.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.