Cutie

Why we included this project

Cutie tracks a single object across a video clip once you supply the initial mask, which saves a lot of tedious frame-by-frame clicking. The scripting demo is the fastest way in: you hand it the first frame's mask, loop through the remaining frames in a plain PyTorch loop, and read back a per-object probability map at each step. That fits editing workflows like rotoscoping or pulling a subject out before compositing, and for people who would rather click than code, the bundled GUI also brings in the permanent memory option from XMem++ for tighter control over how objects are remembered. It's the follow-up to XMem with better consistency and speed, and the training and evaluation docs cover reproducing results on standard benchmarks if you're weighing segmentation models.