hrgentry/TangoSeg-Data-Splits

Materials of the revised manuscript, in a new revision_2026/ directory. Every file from v1.8.0 outside that directory is byte-identical. The revision retrains every matrix cell with three seeds and scores it at a threshold selected on held-out data, repeats the partition controls with fixed test images, separates training duration from schedule rescaling, extends the matched TANGO contrasts to five seeds with crack-geometry metrics, compares the two pretraining sources against one shared reference, and retrains six models at native resolution. This release carries what those results are computed from: machine-readable results for every revised table and section; per-image score histograms of every evaluated checkpoint (1,001 bins of foreground and background pixel counts per image), from which every threshold rule, F1, average precision and bootstrap interval in the paper can be recomputed; per-image crack-geometry metrics at the held-out threshold; the new split index files: three fixed-test Crack500 partitions and the CrackMap five-fold partitions; per-epoch validation scalars and the full dumped configuration of each of 859 training runs (seeds are in the configurations); the training configurations, evaluation and statistics scripts; the revised supplementary material. No dataset images or annotations are redistributed, and no model weights or third-party model source code are included. revision_2026/README.md maps each file to the manuscript section it supports.