Image-mask-spurious triplets dataset for geometry-aware road extraction from UAV imagery

This article presents a dataset of 3309 image-mask-prediction triplets derived from UAV orthophotography over six municipalities in Antioquia, Atlántico, and Chocó, Colombia, where predicted segmentation outputs exhibit degraded geometric quality such as noisy boundaries, fragmented regions, and angular distortions relative to the reference geometry. The dataset provides paired instances of geometrically regular ground-truth masks and degraded segmentation predictions, enabling the development of post-processing algorithms for boundary regularization and road vectorization. Non-overlapping 512 × 512-pixel RGB tiles were extracted at a stride of 512 pixels, retaining only tiles with a minimum road coverage of 10%. The dataset is partitioned into training (2084 samples; 63.0%), validation (232 samples; 7.0%), and test (993 samples; 30.0%) subsets. Each sample consists of three co-registered files: a raw UAV image tile, a binary ground-truth mask produced through manual digitization, and a predicted mask generated by the best-performing architecture from a five-model benchmark (SegFormer-B3). Connected-component analysis identified 3562 discrete road segment instances, and integrity validation confirmed zero-dimension mismatches across all 3309 pairs. The predicted masks exhibit boundary irregularities, width distortions, and angular deviations whose explicit pairing with ground-truth references enables quantitative evaluation of geometry-aware refinement methods.