// the find
wgcban/ChangeFormer
[IGARSS'22]: A Transformer-Based Siamese Network for Change Detection
ChangeFormer is the reference implementation for a 2022 IGARSS paper that does binary change detection on paired satellite images using a transformer-based Siamese network (a SegFormer-style encoder feeding a difference module). It's aimed at remote sensing researchers who want to reproduce or benchmark against the LEVIR-CD/DSIFN-CD results, not at anyone looking for a drop-in production change-detection tool.
Ships working end-to-end demos with sample image pairs and pretrained checkpoints for both LEVIR-CD and DSIFN-CD, so you can run inference in minutes instead of assembling a training pipeline first. It also bundles several baseline models (BIT, DTCDSCN, SiamUnet_conc/diff, plain Unet) with matching train/eval scripts, making it useful as a benchmark harness rather than just a single model. The SegFormer/ADE20k pretraining step for faster convergence is documented with the actual checkpoint link, not just mentioned in the paper.
License is research/non-commercial only, so anyone wanting to use it in a product has to contact the authors first. The shell scripts hardcode absolute paths like /media/lidan/ssd2/ChangeFormer/checkpoints, so every run requires manually editing scripts rather than passing a config or env var. Dependencies are pinned to Python 3.8 / PyTorch 1.10.1 / torchvision 0.11.2 with no newer compatibility notes, and the repo has had no commits since January 2024, so expect friction on a modern CUDA/driver stack. Multi-class change detection isn't a supported mode — the README points to a GitHub issue comment describing four separate manual code edits across different files to make it work, which is a maintenance smell for anything beyond the binary LEVIR/DSIFN use case.