From Classical Pipelines To Promptable Foundation Models: A Cross-Domain Survey Of Thin-Object Segmentation For Power Lines, Cracks, And Retinal Vessels

Document Type

Article

Publication Date

1-1-2026

School

Computing Sciences and Computer Engineering

Abstract

Thin-object segmentation plays a key role in accurately identifying elongated and topology-sensitive structures such as power lines, pavement cracks, and retinal vessels. These structures are central to many IoT-enabled monitoring systems, where data are collected by distributed cameras, unmanned aerial vehicles, and biomedical sensors. Despite steady advances in computer vision, reliably segmenting thin and discontinuous objects remains difficult due to class imbalance, low contrast, and frequent fragmentation. While the underlying applications differ, these domains share common geometric and computational challenges, motivating a unified analysis. This paper presents a comprehensive survey of thin-object segmentation methods across industrial, civil, and biomedical imagery. We introduce a method-centric taxonomy that organizes prior work into five categories: CNN-based architectures, Transformer and hybrid models, GAN and generative approaches, classical image-processing methods, and foundation or promptable models. In addition, we consolidate widely used public datasets from the three domains to support consistent cross-domain comparison. Beyond the survey, we provide a focused analytical evaluation of recent foundation and vision–language models to illustrate their behavior on thin-object segmentation tasks. The analysis highlight recurring difficulties in preserving ultra-thin structures and reveal limitations of conventional pixel-level metrics in reflecting perceptual and structural quality. By combining taxonomy, dataset consolidation, and targeted experimentation, this work offers a unified perspective on thin-object segmentation and clarifies the opportunities and challenges of applying promptable foundation models to this problem.

Publication Title

IEEE Internet of Things Journal

Link to Full Text

Share

COinS