FaultSeg: A Dataset for Train Wheel Defect Detection

The dataset contains original raw images of train wheels captured using a GoPro Hero 9 Black camera, along with their respective segmentation labels for real-time wheel defect detection. The images are annotated for four distinct classes: Wheel, Shelling, Discoloration, and Cracks/Scratches. It is pertinent to mention here that the model confuses between following classes: peeling, cracking, and scratches. We have categorised all of the cracks and scratches in our dataset into a single class called cracks/scratches.   Annotated Data:This data is further divided into formats and stored within three folders: train, test, and valid. The formats include: — JSON: Located in the “Labeled_data_coco_segmentation_JSON.zip” folder.— XML: Found in the “Labeled_data_voc_XML.zip” folder.— TXT: Available in the “Labeled_data_TXT.zip” folder.— TFRecord: Under the “Labeled_data_tfrecord.zip” folder.— CSV: Located in the “labeled_data_multiclass_CSV.zip” folder. These formats strengthen the overall usability of the code by facilitating the training of various AI-based models, including YOLO, Detectron 2, FastInst, and many others.   For detailed annotation of the dataset, please go through this Roboflow link.