RoadPaint benchmark: map-referenced, human-verified vehicle holes in aerial orthoimagery

Benchmark data, code and model weights accompanying the article “Inpainting Vehicle Holes in Aerial Orthoimagery: A Map-Referenced, Human-Verified Benchmark and a Study of Road Priors” (D. Park, W. Kim and H. Ahn), submitted to IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing. roadpaint_release_v1.zip contains the benchmark: hole masks of 600 single-vehicle, 137 fragmented and 139 vehicle-queue holes on 0.24-m orthophoto tiles of Seoul, Incheon and Suwon, and 536 held-out holes in Hwaseong; tile identifiers; the registered origins of the NGII orthophoto sheets; the human verification and lane-marking annotations; per-hole scores of every configuration in the article, on the tiles used and on the released renders; and the code to render the tiles, score a method and run the hole-content invariance test. roadpaint_weights_v1.zip contains the first training runs of the fine-tuned LaMa, the fine-tuned ZITS++ and the rotation-trained RoadPaint. No imagery is included. The tiles are rendered from the orthophotos of the National Geographic Information Institute (NGII) of the Republic of Korea with the included script; see README.md. Data and annotations: CC BY 4.0. Code: MIT. The fine-tuned LaMa and ZITS++ weights remain subject to the licenses of the original weights and code.