facebookresearch/CutLER ? reverse-engineered prompt

Build me a working version of this project for unsupervised object detection and instance segmentation.

I want a setup where I can generate pseudo masks from unlabeled images with MaskCut, then use those masks to train a CutLER model and run inference on new images. Please include a simple demo flow for local testing, plus a way to run on ImageNet style data and save the generated annotations. If the repo also supports the video version, wire that in too so I can try VideoCutLER on a sample video or frames.

Make it easy to install, easy to run a demo, and easy to swap in my own images or dataset paths. Add clear commands or scripts for the main steps, and sensible defaults so I can get something working quickly without needing to understand the full paper first. If you need to check current setup details or dependencies, look up the latest docs online.