CascadePSP

Why we included this project

Segmentation models often produce masks that look clean until you zoom in, when boundary detail starts to blur. CascadePSP tackles that by refining an existing binary mask against the original image: a global pass keeps the overall structure while a local pass works on high-resolution patches to tighten edges. It is class-agnostic, so it works on top of whatever segmenter or matting model you already have. The repository includes a pip package that refines an image in two lines of code, plus a pretrained model, training code, and an annotated UHD dataset called BIG for anyone who wants to retrain or dig deeper. For teams upgrading output quality without retraining their main model, it is a low-effort way to sharpen mask boundaries.