I used it for segmentation for this home climbing/spray wall project: https://freeclimbs.org/wall/demo/edit-set https://freeclimbs.org/wall/demo/edit-set It do

I used it for segmentation for this home climbing/spray wall project: https://freeclimbs.org/wall/demo/edit-set https://freeclimbs.org/wall/demo/edit-set It does detection on the backend and then feeds those bounding boxes into SAM running in the browser. This is a little slow on the first pass but allows the user the adjust the bboxes and get new segmentations in nearly real time, without putting a ton of load on the server. Saved me having to label a bunch of holds with precise masks/polygons (I labeled 10k for the detection model and that was quite enough). I might try using SAM's output to train a smaller model in the future, haven't gotten around to it. (Site is early in development and not ready for actual users, but feel free to mess around.)