I just open-sourced (MIT License) semantic segmentation annotation powered by segment-anything model that I used for a while in my projects. Hopefully it will help someone.
Link to my project: SAMAT (any feedback in Discussions section on GitHub is appreciated)
Features:
- Brush annotation (opposed to polygons)
- Magic Wand (like in Photoshop) powered by segment-anything masks (it is optional btw, if you don’t have cool GPU to prepare masks)
Why yet another annotation tool?
Before starting this project I tried supervisely, segments.ai, roboflow and several others, but found them not suitable enough for my projects.
Their cons, I tried to fix with my tool:
- Latency: they are web based, hence burden with irritating latency during labeling (SAMAT is snappy because it is local desktop app)
- Complicated: too much features, hence overloaded UI (SAMAT is just a colored brush)
- Proprietary: they all follow SaaS model
Btw, there is labeling tool called SALT on github which also uses segment-anything model, but have a different approach to UI/UX, so take a look at it too.