MuraTrace: A Real-Background Synthetic Benchmark for Multi-Scale Display Mura Inspection (v1.0)

MuraTrace is a real-background synthetic benchmark for multi-scale display Mura inspection with paired perceptual and process-trace annotations (IEEE/CVF WACV 2027, Datasets track). Physically motivated Mura defects (multiplicative luminance fields with SEMU-controlled visibility) are composited onto 30 defect-free real mobile panel backgrounds. The package contains 540 synthetic glasses (8x7 panels each), 30,240 panel images (16-bit PNG, 727x359), and 5,736 unique visible defect instances in five classes (spot, line, region, gradient, band) and eight subtypes. Every instance carries a dense semantic mask, an instance map, a bounding box, subtype, polarity, panel-instance SEMU visibility, two reject labels (author thresholds and expert-consensus thresholds), and process-trace metadata (process step, tool, fault mode, time window, route-graph node). Every glass carries a 21-node route graph in two configurations (default and ambiguous-cause). The package ships the background-level split used in the paper (18/6/6 real backgrounds for train/val/test), a glass-disjoint split protocol, the generation configuration and SEMU thresholds, the perceptual-study instrument with anonymized responses, and SHA256 checksums. These files are the exact data on which all reported numbers were computed. Code (generation, training, evaluation, analysis) and baseline checkpoints: https://github.com/su9yeong/MuraTrace. Unpack every zip part into the same folder and verify with "python tools/verify_release.py --root <folder> --sha". Citation: S. Lee and D.-H. Lee, "MuraTrace: A Real-Background Synthetic Benchmark for Multi-Scale Display Mura Inspection," IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2027.