Onboard_RSCD_data — 256×256 change-detection pools (LEVIR / SYSU / WHU)

These are the exact data pools used for the teacher and student experiments in kimongjin/Onboard_RSCD (weights and Baseline/ code). The training scripts read them directly with --data_root <DATA_ROOT>/<pool>.

Files

File	Pool	train / val / test	Size
LEVIR_256_pool.tar	LEVIR-CD	6096 / 1024 / 2048	~2.5 GB
SYSU_256_pool.tar	SYSU-CD	12000 / 4000 / 4000	~5.0 GB
WHU_CD_256_pool.tar	WHU-CD (rebuilt 2026-09-25)	4354 / 484 / 2596	~1.9 GB
MD5SUMS	md5 of the three tars

hf download kimongjin/Onboard_RSCD_data --repo-type dataset --local-dir $DATA_ROOT
cd $DATA_ROOT && md5sum -c MD5SUMS && for t in *.tar; do tar xf $t; done

Each pool has the same layout: A/ (time 1), B/ (time 2), label/ (0/255 PNG), and list/{train,val,test}.txt. File names carry a split prefix (train_*, val_*, test_*), so all splits share one flat folder.

How each pool was made

• LEVIR-CD: the 1024×1024 originals are cut into sixteen non-overlapping 256 tiles.
  • Our raw copy lacks the official val set (64 images). We split 64 of the 445 train images off as val, so val tiles have train_ source IDs.
  • Test is the official 128 images.
  • Labels match crops of the original change labels exactly, and no source image appears in two splits.
• SYSU-CD: identical to the original 256×256 release. Only the split prefix is added.
• WHU-CD: built from the official whole images.
  • A = 2012, B = 2016, label = change_label/{train,test}/change_label.tif, all cut into non-overlapping 256 tiles from the top-left. The trailing remainder is dropped.
  • Val is a random 10% of the train tiles (seed 42).
  • The tiles re-assemble to the official change label with 100% pixel agreement. Positive pixels: train 4.55%, val 4.21%, test 3.68%.
  • Script: Baseline/data_prep/prep_whu_cd.py in the weights repo.

⚠️ An older pool, WHU_256_pool (4536/504/2760), used the 2012 building mask instead of the change label. It is not included here. Results obtained on it are not change detection.

License

These pools are derived from the public LEVIR-CD, SYSU-CD and WHU Building CD datasets. They are for internal research use only, under each original dataset's terms. Please cite the original datasets.