Weather Adverse Road Defect Semantic Segmentation

OverviewThis dataset is an inference and robustness benchmark designed for a Master's thesis evaluating models like Mask2Former and MMSegmentation. It features anonymized GoPro video frames of road defects captured across three environmental moisture levels: dry, wet, and half (mixed).

Dataset StructureThe dataset follows a flat directory structure to make it instantly loadable with basic PyTorch/TensorFlow data loaders:- images: Anonymized road surface images (JPEGs/PNGs).- labels: Grayscale single-channel semantic masks (Class IDs).- instances: Grayscale multi-channel instance masks (Unique Object IDs for counting).- labels_color: 3-channel RGB visual masks for easy human previewing.- metadata.csv: Contains filename, condition (dry/wet/half), image dimensions, and sequence_id (original video folder name to prevent data leakage during testing).

Classes- 0: bg - 1: cracks - 2: cracks_alligator- 3: cracks_severe- 4: edge_cracks- 5: fretting- 6: pothole- 7: manhole- 8: pole_shadow