Hi-BCD Amsterdam (train) and Rotterdam (dev) public cases: 256x256 DSM-to-image building-change crops

# Hi-BCD Amsterdam (train) and Rotterdam (dev) public cases — mirror This record mirrors the public training/development cases used by a reproducibility task built on the Hi-BCD dataset: - Liu, Biyuan; Chen, Huaixin; Li, Kun; Yang, Michael Ying. "Transformer-based Multimodal Change Detection with Multitask Consistency Constraints", Information Fusion 2024. arXiv:2310.09276. - Source code and dataset links: https://github.com/HATFormer/MMCD ## Contents - `hibcd_public_train.tar` — 1,212 labelled Amsterdam cases (`train/.npz`) plus `manifest_train.csv`. SHA-256: 9a43b3da1008483b38de642d87134b4b7e3962ca2f84e87554e575b3c9cafe27 Size: 327567360 bytes. - `hibcd_public_dev.tar` — 1,061 labelled Rotterdam cases (`dev/.npz`) plus `manifest_dev.csv`. SHA-256: 0d4f4be0c627a0b2b8f73b8d21fd66a941371e417a74bcb45032d33df7455e9e Size: 268175360 bytes. Each `.npz` contains `img` (uint8, 256x256x3, post-event aerial RGB), `dsm` (float32, 256x256, pre-event surface height in metres at 0.25 m) and `gt` (uint8, 256x256, values in {0,1,2}: 0 unchanged, 1 demolished, 2 newly-built). Manifest columns are `case_id,city,source_tile,crop_y,crop_x,demolished_pixels,newly_built_pixels`. Manifest SHA-256: - manifest_train.csv: 02b115db12675d01a3a4e9a91af38bff938d6f1cde36478362002d64aea0ebc2 - manifest_dev.csv: eba98e70f1657f986e453b43a78221b710908374e31440433561f0e0a291c241 Combined per-file SHA-256 (SHA-256 over the concatenation, in sorted case_id order, of each case file name and its binary SHA-256 digest): - train: ef48f4aa79373dd8d35b5dcade3c5ed617ac48bf4bca671fc5253d93a12e0a20 - dev: a55b767c0f314b7b6e30115d4e991d085a9b08f7f650dcc1e9608d72e9b99feb Ground-truth pixel totals: train 1,871,529 class-1 and 5,068,188 class-2 pixels (499 zero-change cases, 500 source tiles); dev 1,286,351 class-1 and 4,078,056 class-2 pixels (500 zero-change cases, 500 source tiles). ## Provenance and generation protocol The cases were generated from the official Hi-BCD source archives, obtained from the dataset's Google Drive folder, with SHA-256 verified before use: - amsterdam_025.zip: 99915a74666921c5c6fe286baf392b742afc7c93cb36f6de15ace0a46cb3b642 - rotterdam_025.zip: c2cb986eaf1b50bc40c7dbb3d4fe2903a689e694a999a8d7dfb3fe9017bac37e Protocol (identical to the task's public generation): each 1000x1000 source tile (all 500 tiles per city, source train/test/val assignments pooled within the city) is cut into native 256x256 crops with upper-left origins in {0, 248, 496, 744} in both axes, without resizing. A crop is included when it contains at least 100 pixels with ground-truth values 1 or 2 in total; the first crop with zero changed pixels, in increasing crop_y then crop_x order, is also included per source tile when such a crop exists. Source label value 3 is mapped to class 2, source label values 0 and 1 are retained, only the first three source image channels are retained, and any DSM pixel that is non-finite or outside [-100, 1000] metres is replaced by the median of the valid DSM pixels in that crop. No Utrecht tile or crop is included in this record; Utrecht is reserved as a held-out evaluation city in the task that uses these public cases. This mirror exists because the source Google Drive files are subject to download-quota errors ("Too many users have viewed or downloaded this file recently") that made the original Drive-based setup fail on fresh machines. The original Hi-BCD terms apply to the data; this record only re-hosts the derived public cases for reproducible, non-commercial research use, with attribution to the original authors.