BROWSE
Detail
Journal Article
Zero-Shot Anomaly Segmentation via Query-Level Uncertainty Analysis
- Authors
- Kimin Yun, Yuseok Bae
- Issue Date
- 2025-12
- Citation
- IEEE Access, v.13, pp.215519-215532
- ISSN
- 2169-3536
- Publisher
- IEEE
- Language
- English
- Type
- Journal Article
- Abstract
- We propose a training-free method for anomaly segmentation that leverages the internal signals of pretrained transformer-based segmentation models. Unlike prior approaches that rely on retraining, auxiliary outlier data, or external supervision, our method identifies out-of-distribution (OOD) regions by analyzing query-level behaviors. Specifically, we suppress confidently predicted inliers while amplifying uncertain predictions through entropy-guided selection and spatial consistency analysis. To improve spatial coherence, we introduce an object-aware refinement step that enforces consistency within class-agnostic segments, yielding anomaly maps that align more closely with object boundaries. Our approach operates in a strictly zero-shot setting and requires no modification of the base model. Experiments across standard benchmarks, including SMIYC, Fishyscapes, and RoadAnomaly, demonstrate that the proposed method achieves state-of-the-art performance among zero-shot approaches and narrows the gap to training-based methods, while providing interpretable insights into model uncertainty. This competency-driven framework demonstrates that robust anomaly segmentation can be achieved without retraining or external data, making it suitable for practical real-world deployment.
- Keyword
- anomaly segmentation, model competency analysis, object-aware refinement, Out-of-distribution detection, transformer-based segmentation
- KSP Keywords
- Art performance, Consistency analysis, Model uncertainty, Object-aware, Real-world deployment, Spatial coherence, Spatial consistency, Zero-shot, outlier data, state-of-The-Art, transformer-based
This work is distributed under the term of Creative Commons License (CCL)
(CC BY)
(CC BY)
