Does SAM 3.1 support SAM 1’s interactive image segmentation task (point/box-based)?

[pank98-hash]

Great thanks for simplifying the original Meta SAM series modules. I mainly focus on SAM3.1 interactive segmentation function. I loaded SAM3.1 and ran simple tests, but found its interactive segmentation ability dropped obviously with unsatisfactory results. Have you ever met this issue?

[heyoeyo]

From what I've seen, v3 and v3.1 tend to give similar masks. The one exception is for objects that take up a lot of the image, where v3.1 can be a bit worse than v3.0. I imagine there's also specific imagery (e.g. satellite, medical, text etc.) that may do better on v3 than v3.1 and vice versa.

On the other hand, v3.1 gives much better IoU predictions, which makes it more useful for automated mask selection when doing image segmentation. In fact, I think v3 makes the worst IoU predictions out of any (v1/v2/v3) of the SAM models, so v3.1 is a big improvement in that sense.

That being said, the v3 update seems to be more about adding features (object detection and multiplexing), rather than improving segmentation quality. So the v1/v2 models are just as good (imo) and worth checking out if you haven't already. There's also sam-hq which does a better job on intricate details.