SutaSet: a multi-label image dataset for classification and object detection for threads

Abstract

SutaSet is a publicly oriented multi-label image dataset designed for fine-grained visual inspection of individual threads on fabric surfaces. It contains 6,699 unique 512×512 PNG images with oriented bounding-box annotations covering four visible thread conditions: normal, frayed, snagged, and visually taut, while images without target threads represent the derived absence condition. Capture-grouped train, validation, and test splits ensure evaluation integrity, with classification and oriented-detection benchmarks provided using ConvNeXt-Tiny, Swin-Tiny, and YOLO11n-OBB. The dataset supports standardized textile defect classification and rotated-region localization of thin thread structures.