Leakage-Controlled Road Damage Classification Dataset and Code

This release supports the reproducibility of the companion paper's Stage 1 (image classification) result: macro-F1 = 0.9528 on a leakage-controlled, locked test set, across four road damage classes (D00: longitudinal crack, D10: transverse crack, D20: alligator crack, D40: pothole). Contents:- 21,041 cropped road-damage images (one folder per class), extracted from 94 driving sessions and letterbox-resized to preserve their original aspect ratio- manifest_with_split_v3.csv: full manifest with each image's session, frame ID, damage class, and its assigned split (train/validation/test) under a leakage-controlled protocol that checks for session-level, temporal, and spatial data overlap- Trained model weights (PyTorch) for the final classifier: a partially fine-tuned YOLO11s-OBB backbone plus a small classification head- Python code to rebuild the leakage-controlled split, train the model, and evaluate it on the locked test set How the labels were produced: the labels are NOT independent of a detector. They were produced in CVAT with a model-assisted workflow. The YOLO11s-OBB detector that is also used as the classification backbone (pretrained on IRDD) first generated boxes and classes automatically, and the author then reviewed every frame and corrected the annotations that were wrong. The automatic annotation files were not kept, so the number of corrections is unknown, and no second annotator took part. The companion paper reports how close the labels are to the detector's own predictions and how the classifier performs on labels with and without a detector counterpart. Not included: raw driving-session videos and precise GPS trajectories (privacy), the sensor time-series data used in the sensor-fusion experiments of the companion paper, and the pretrained backbone checkpoint. The classification backbone (YOLO11s-OBB) was pretrained separately on the IRDD dataset (Taher, 2026, https://doi.org/10.5281/zenodo.21167531), which contains no overlap with the images released here. Version history: version 2 corrected the model name (YOLO11s-OBB, not YOLOv11n-OBB). Version 3 corrects the description of how the labels were produced in this description and in README.md. The data files are unchanged. See README.md inside the archive for the folder structure and reproduction steps.