FSC-180k
We introduce our hybrid semantic change detection dataset, named FSC-180k. It consists of approximately 60,000 real aerial images sourced from the FLAIR dataset, along with 180,000 artificially modified images. These images were generated using our HySCDG pipeline applied (three times) to each real image.
In total, the dataset provides 180,000 image pairs. Each pair is accompanied by a binary change map and semantic segmentation maps for both images (land use classification using the 19 classes of FLAIR).
Dataset Structure
The folder IMG contains the 60,000 real images, while each of the folders IMG11, IMG12, IMG13 contain their corresponding artificially modified pairs.
Inside each folder, images are named following the FLAIR nomenclature : IMG_XXXXXX.tif (with XXXXXX from 000001 to 061712). And splitted into subfolders of 1,000 images : 00, 01, ..., 61.
The structure is exactly the same for the labels (binary change and semantic maps). They are named LBL_XXXXXX.tif and stored in the folders LBL11, LBL12, LBL13, matching exactly the images organization. Note that there is one label file per image pair (and not per image).
Each file (images and labels) is in TIFF format.
Original images (folder IMG) are composed of 5 bands (R,G,B,Infrared,Elevation)
Generated images contain 3 bands (R,G,B).
Labels contain 3 bands :
- binary change map on the first layer (0:no change, 255:change)
- semantic map of first image (real image) on the second layer (classes between 0 and 19)
- semantic map of second image (generated image) on the third layer (classes between 0 and 19)