WHU_Generative_Change_Detection_Dataset

Click Here to download.


1.Training Set

-- gcd: Generative change samples synthesized from three semantic segmentation datasets (3SS).

- im1/: Original images from 3SS.

- im2/: Generated 'changed' images (for each x.png in im1/, multiple x_i.png may exist in im2/, where the i-th class is altered: 1-background, 2-building, 3-road, 4-water, 5-barren, 6-forest, 7-agriculture, 0-no-data)

- label/: Binary change labels (0 or 255)

- mask1/ & mask2/: Semantic masks for t1 and t2 (values in 0-6, 0-background, 1-building, 2-road, 3-water, 4-barren, 5-forest, 6-agriculture)

-- Total of 21,000 paired change samples.


-- ugcd_full: Change and negative samples collected from Google Earth.

- Same subfolders as gcd/, but mask1/ and mask2/ contain only zero values (ignored during semantic change detection training).

-- Total of 3,167 pairs.


-- ucd: One sub set of ugcd_full. Pure negative samples (without generative change pixels).

- Same folder structure as above.

-- Total of 2,089 pairs.


-- ugcd: The other sub set of ugcd_full.Generative samples (including all generative change samples based on SAM and SD).

- Same folder structure as above.

-- Total of 1,078 pairs.


Training recommendation:

- If using gcd + ugcd_full: concatenate at a 3:1 ratio, and sample according to that ratio in each batch.

- If using gcd + ucd + ugcd: concatenate in a 6:1:1 ratio, and sample according to that ratio in each batch.

--------------------------------------

2.Validation Set

- Follows the same folder structure as train/ (no subdivision into gcd/ucd/etc.).

- Contains im1/, im2/, label/, mask1/, mask2/.

-- Total of 600 pairs.

--------------------------------------

3.Test Set

-- test/ (generative):

- Supports both Binary Change Detection (BCD) and Semantic Change Detection (SCD).

- Follows the same folder structure as val/.

-- Total of 3,300 pairs.


-- test2/ (real-world):

- Supports BCD only.

- Contains im1/, im2/, label/ (no semantic masks).

-- Total of 3,906 pairs.

--------------------------------------

4.Directory Structure

WHU-GCD

|-train/

| |-gcd/ # Generative change samples from SS3

| | |-im1/ # Original images

| | |-im2/ # Generated images

| | |-label/ # Binary labels (0, 255)

| | |-mask1/ # Semantic masks t1 (0-6)

| | |-mask2/ # Semantic masks t2 (0-6)

| |-ugcd_full/ # Google Earth change + negative samples

| | |-im1/

| | |-im2/

| | |-label/

| | |-mask1/ # All zeros (ignored for SCD)

| | |-mask2/

| |-ucd/ # Pure negative samples, subset of ugcd_full

| | |-im1/

| | |-im2/

| | |-label/

| | |-mask1/

| | |-mask2/

| |-ugcd/ # Generative samples, subset of ugcd_full

| | |-im1/

| | |-im2/

| | |-label/

| | |-mask1/

| | |-mask2/

|-val/ # Validation set

| |-im1/

| |-im2/

| |-label/

| |-mask1/

| |-mask2/

|-test/ # Generative test set (BCD + SCD)

| |-im1/

| |-im2/

| |-label/

| |-mask1/

| |-mask2/

|-test2/ # Real-world test set (BCD only)

| |-im1/

| |-im2/

| |-label/

|-README.md/