NA-CD Dataset (Non-Aligned Change Detection Dataset)

Synthetic change regions are introduced by corruption and perturbation, and the pre-change imagery is generated by applying HSV and affine perturbations to the post-change tiles.</li> <li><b>LEVIR-CD_train.zip / LEVIR-CD_test.zip / LEVIR-CD_val.zip</b> — 1280×1280 tiles from LEVIR-CD. <code>images/</code> holds the post-change tiles, <code>images0/</code> the pre-change tiles, and <code>labels/</code> the change masks (<code>.npy</code>) together with instance polygons (<code>.pol</code>). <code>names.txt</code> lists the tile names.</li> <li><b>WHU-BCD.zip</b> — Tiles from WHU-BCD with the same <code>images/</code>, <code>images0/</code>, <code>labels/</code> layout, plus <code>A23_0/</code> and <code>A23_1/</code> containing the ground-truth affine transformation matrices A₂₃ used to generate the non-aligned pairs, and <code>pol_0/</code>, <code>pol_1/</code> with the polygon annotations of each temporal instance.</li> <li><b>Google_patches_seg.zip</b> — Tiled training and validation data from the Google (CD_Data_GZ) dataset: <code>train/patches_seg/</code> and <code>val/patches_seg/</code>, each with <code>images/</code> (post-change), <code>images0/</code> (pre-change) and <code>labels/</code>.</li> <li><b>Google_original_images.zip</b> — The original large Google Earth images (<code>images/</code>, <code>images0/</code>) and their label maps (<code>label/</code>) before tiling, provided for provenance.</li> </ul> <p dir="ltr">All non-aligned pairs are generated with controlled affine transformations (translation, rotation, scaling); the ground-truth A₂₃ matrices are released so that geometric-correction accuracy (SC) can be evaluated.</p>