pytorch-semseg

Why we included this project

For anyone working on semantic segmentation, this repo collects the classic architectures in one PyTorch codebase: FCN, U-Net, SegNet, PSPNet, ICNet, LinkNet, and FRRN. Because they all share the same training loop, you can run them on the same data and see how the different designs actually compare, rather than assembling seven separate projects with their own conventions. The bundled data loaders cover CamVid, Cityscapes, Pascal VOC, ADE20K, and NYU Depth, so starting real benchmark training takes minimal setup. The models are implemented directly in PyTorch, not wrapped behind a heavyweight framework, which keeps the code readable and easy to adapt when you want to add your own dataset or tweak an architecture. It leans toward a learning and reference toolkit rather than a maintained production service, but for research, coursework, or a quick prototype, it is a convenient starting point with everything in one place.