Change detection's mixtape: every model, one repo

What it does

Open-CD is a Python toolbox for remote-sensing change detection. It wraps implementations of roughly twenty published models—including Changer, BiT, and ChangeFormer—into a shared training and evaluation framework built atop OpenMMLab and mmsegmentation. It also bundles standard datasets such as LEVIR-CD and WHU-CD, aiming to turn model-hopping from a porting exercise into a config swap.

The interesting bit

The project doesn’t hide its integrator nature; it openly bills itself as “standing on the shoulders of giants” and borrows code directly from original repositories. The real value is standardization: it forces a fragmented research landscape into one consistent API, complete with a technical report accepted at ACMMM 2025.

Key highlights

• Supports ~20 change detection architectures, from classic siamese networks to recent transformers.
• Built on OpenMMLab 2.0 and PyTorch 2.0, inheriting their training engines and data pipelines.
• Bundles over ten public satellite datasets, including LEVIR-CD, WHU-CD, and SECOND.
• Published technical report (ACMMM 2025) and a HuggingFace model zoo.
• Provides an inference API and a Colab demo for quick experiments.

Verdict

Researchers working on remote-sensing change detection who are tired of juggling one-off model repositories will find this a convenient consolidation layer. If you want a ground-up framework with no external dependencies, this is essentially curated glue code around OpenMMLab.