Using box + point gives worse results comparing to only boxes

Hi,

I have an object detection model and I want to create masks for those objects using SAM, I have done prompting with the output boxes given by my obj detection model but results given by SAM was not as I wanted (results are good but there are some errors), so I decided to add a point with box (the point is the center of the box which is on the object for almost all images), but results given now are a little worse than the first try (for images in which the point is clearly on the object).

I have an object detection model and I want to create masks for these objects using SAM, I have done prompts with the output boxes given by my obj detection model but the results given by SAM were not as I wanted (the results are good but there are some errors), so I decided to add a point with the box (the point is the box center which is on the object for almost all images), but the results given now are a little worse than the first try (for images for which the point is clearly on the object)..

Any explication for that?

Thanks,

The report describes worse masks when combining object-detection boxes with center-point prompts in SAM, but names no repository file, test, or reproducible example. Start by reproducing the comparison with box-only and box-plus-point prompts, then determine whether the behavior is expected or represents a fixable defect.

Written by the indexing model from the issue text.