Segment Anything 2

[Acly]

https://github.com/facebookresearch/segment-anything-2

• It's fast and can be quite small.
• Detection should be improved
• Similar to first model it runs on fixed resolution (512 by default)
• Binary masks

So for precise masks some kind of follow up would still be needed.

No ONNX model out yet.

[Rubenacp]

Would be great

[Acly]

Did some tests with exported ONNX models. They're working fine, but performance wise it would be a big regression.

Model	Size	Inference time
MobileSAM	43 MB	450 ms
SAM2 Small	179 MB	1240 ms
SAM2 Tiny	151 MB	1000 ms

This is on CPU, which is most common scenario. The masks are mostly the same regarding pixel quality. The detection accuracy is expected to better for SAM2.

A distilled/quantized version of SAM2 model would be interesting...

[Rubenacp]

Thank you, I 'll try it

[feffy380]

Note: Meta has released updated models (SAM 2.1)