SMARTEX Textile Hyperspectral Classification Dataset

SMARTEX dataset provides RGB and hyperspectral (HSI) scans of various textile materials with the goal of enabling pixel-wise material classification and segmentation. The garments represented are primarily composed of either pure materials or binary material composites—no item contains more than two textile materials. Dataset Contents Each data sample includes the following components: RGB ImageCaptured using a line-scanning Alkeria NECTA N2K2-7C camera.These images are not spatially aligned with the hyperspectral cubes and are provided exclusively for visual reference during annotation. Hyperspectral CubeAcquired using an InnoSpec RedEye 1.7 NIR line-scanning hyperspectral sensor.Each cube contains 252 spectral channels in the 900–1700 nm wavelength range, stored as .npy files with shape (H, W, 252). Files are split into 10 archives each containing 12 hsi cube Stacked ImageVisualization aid combining: One selected HSI channel (used for annotation), The corresponding RGB image (to guide labelers). These images are generated to support manual labeling in CVAT, allowing annotators to work on HSI-derived content while using RGB for additional visual context. Acquisition Images were acquired using a conveyor belt (to allow 2D acquisition for the line-scanning RGB and hyperspectral cameras), and managed using a ROS-based system developed for acquiring this dataset. The data collector selects lines for each image feed and assigns One label for each garment, which represents the material composition (as percentages of present materials). This initial labels are available in labels/smartex_dataset_compositions.jsonl Sample Types The dataset includes two distinct categories of textile items: Fabric Swatches – clean, flat textile samples obtained from known material sources. Everyday Garments – real clothing items collected and scanned directly by the research team. These two sample types help balance controlled material conditions with realistic visual complexity. Annotations Verified annotations are provided in the COCO format (labels/smartex_annotations_cocostyle.json) with pixel-wise segmentation masks for the following three semantic classes: Background – any area not belonging to textile or non-textile (essentially the conveyor belt). Non-textile Material – elements like zippers, buttons, etc. Textile – the primary foreground class, representing garment fabric. In addition, each textile mask is enriched with material composition attributes, specifying the percentage distribution of up to two materials from the following list: Cotton Wool Polyester Acrylic Elastan Viscose Nylon These attributes are stored in the attributes field of each COCO annotation as a dictionary of {material_name: percentage}. In addition, we provide class-wise binary mask for each image (in masks folder) which are directly derived from the annotations file. Metadata & Structure Each sample is identified by a unique UUID-based name. An optional sample identifier or description is stored in the flickr_url field of the COCO images section. This can be used to link the sample to a catalog ID or fabric source. Acknowledgement This project was funded (as an open call beneficary) by euROBIN EU Project (GA 101070596)