TexSortMLDataset

Le résumé fourni par la source

TexSortML: A Dataset for Garment Classification in Automated Textile Sorting Description Public clothing datasets show garments flat, fully visible and canonically oriented — a presentation that does not survive contact with a real sorting line, where items arrive folded, rotated and self-occluded on a moving conveyor. TexSortML was recorded to close that gap: 751 post-consumer garments across 11 categories, sourced from a professional textile sorting facility in Germany and captured on a moving conveyor belt under three controlled stages of increasing handling disorder. Full documentation of the recording protocol, the category definitions and a five-model baseline evaluation is in the accompanying paper (see Citation). Contents of this version Images — 2,253 labelled frames (751 garments × 3 stages), one frame per garment per stage, labelled with garment category and stage Videos — the raw recordings the frames were extracted from, one video per category and stage Earlier versions are partial: version 1 contains only the images, version 2 only the videos. Categories T-Shirt (56) · Polo Shirt (121) · Shirt (80) · Longsleeve (116) · Jacket / Coat (96) · Jeans (39) · Trousers (41) · Shorts (102) · Dress (55) · Zipper / Sweat Jacket (17) · Top (28) The categories were defined together with the facility's sorting personnel, by observable visual and functional characteristics, and reflect distinctions that matter operationally for reuse and recycling. The distribution is unbalanced, from 17 to 121 items — report class-balanced metrics alongside overall accuracy. Difficulty stages Each garment was recorded across all three stages, which differ in belt speed and placement: Stage Belt speed Placement Stage 1 31 m/min Laid flat and neatly smoothed on the belt Stage 2 62 m/min Placed using one hand, resulting in minor rotation and displacement Stage 3 78 m/min Thrown onto the moving conveyor from a distance of one meter Recordings were made on a motorized conveyor belt with a GoPro camera at 240 fps, inside a light-shielding tent that keeps illumination constant across all recordings. Using the data Split at the garment level, not the frame level. All images of one garment share the same physical item across three stages; a frame-level split places it in both partitions. The paper uses stratified five-fold cross-validation over the 751 garments, stratified by category — reusing that protocol keeps results comparable with the published baselines. For reference, a fine-tuned ViT-large reaches 95.2% / 91.9% / 85.9% mean accuracy on Stages 1–3 under that protocol. Scope. Exactly one garment is in the field of view at a time, so the data captures self-occlusion — folding, crumpling, rotation — but not inter-object occlusion from overlapping garments. Illumination is constant, and all recordings come from a single facility and camera setup. Video files. Individual garments may appear more than once within a video as they move through the conveyor system, and manual positioning corrections are occasionally visible. Both were deliberately preserved as part of what the dataset represents. Citation > Schanz, J.; Kohnle, M.; Kopf, F.; Geldhäuser, S.; Cetin, M.; Teynor, A. TexSortML: A Dataset and Baseline Evaluation for Garment Classification in Automated Textile Sorting. Sensors 2026. [DOI to be added upon publication]

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.