// the find
rishikksh20/ResUnet
Pytorch implementation of ResUnet and ResUnet ++
An unofficial PyTorch implementation of ResUnet and ResUnet++ for image segmentation, built and tested specifically against the Massachusetts Roads Dataset. It's for someone who wants a reference implementation of these two papers to read or adapt, not a drop-in segmentation library.
Both ResUnet and ResUnet++ architectures are implemented from the papers, not just one, and the README links directly to the source papers for each design choice (squeeze-and-excitation, ASPP, CRF+TTA). The repo also includes a working preprocessing pipeline that tiles large satellite images into fixed-size crops, which is the annoying part most reference repos skip.
The author states outright it's hardcoded for one dataset and one directory structure, so the dataloader and preprocessing need to be rewritten for any other use case. Unmaintained since September 2024, and there's no pretrained weights, no benchmark numbers against the papers' reported results, and no inference/demo script — just training code.