IDEA-Research/Grounded-SAM-2

Is there a parameter in the code that can control the tracking accuracy of the video predictor and reduce the false positives (FP)?

Opened this issue · 1 comments

Is there a parameter in the code that can control the tracking accuracy of the video predictor and reduce the false positives (FP)?
I've noticed that when SAM2 loses track of a target, it sometimes predicts a mask that covers the entire image instead.

We did not figure out how to carefully control this in SAM 2, you can refer to SAM 2 issues to see if there're some solutions for this situation.