StitchingNet-Seg, A semantic segmentation dataset for industrial sewing stitch defects

StitchingNet-Seg is a large-scale dataset for semantic segmentation, comprising 10,836 images with pixel-level annotations, designed to advance automated quality inspection in the textile and apparel industry.MotivationAdopting Industry 4.0 technologies in the textile and apparel industry has been historically slow due to the labor-intensive nature of production and the inherent deformability of non-rigid textile materials. While automated quality inspection is well-established for raw fabric defects, sewing stitch defects, which occur dynamically during assembly, still rely heavily on manual visual checks, leading to a defect leakage rate of 5–10% and significant production bottlenecks. Existing datasets often provide only image-level labels or bounding boxes, which are insufficient for the precise geometric analysis needed for high-end quality control. To bridge this gap, we introduce StitchingNet-Seg, a large-scale semantic segmentation dataset comprising 10,836 images with meticulous pixel-level masks across 11 fabric types and 7 defect categories, providing a reliable foundation for developing intelligent Automated Optical Inspection (AOI) systems in the global garment industry.Dataset descriptionThe dataset includes diverse sewing conditions and precise semantic masks to ensure model robustness.Total images: 10,836 images (derived and filtered from the original StitchingNet)Fabric varieties: 11 representative fabric types with various textures and colors: A. Cotton-Poly, B. Linen-Poly, C. Denim-Poly, D. Velveteen-Poly, E. Polyester-Poly, F. Satin-Core, G. Chiffon-Poly, H. Nylon-Core, I. Jacquard-Poly, J. Oxford-Core, and K. Polyester (coated)-CoreThread colors: Combinations of similar and contrasting thread colors.Classes: normal and 7 defective types: Normal, 1. Skipped stitch, 2. Broken stitch, 3. Pinched fabric, 4. Crooked seam, 5. Thread sagging, 7. Stain and damage, and 10. Overlapped stitchResolution: 224 × 224 pixels.Creation detailsOriginal source: StitchingNet (14,565 sewing stitch images)Time period (filtration and annotation): 2025.02 - 2026.01Annotation: Pixel-level semantic masks created using the Computer Vision Annotation Tool (CVAT)Data recordsThe dataset is organized into a hierarchical structure containing various fabric types and sewing defects. For the convenience of the researchers, we provided original images, segmentation masks, and annotation files in the COCO format.Code examplesWe provide reference implementation codes (Python and Google Colab notebook) in the code-examples.zip file to help researchers quickly get started with StitchingNet-Seg.LicenseThe StitchingNet is licensed under CC BY-NC 4.0. This means it is free for research updates and non-commercial use with proper attribution.