The segmentation results are discontinuous and have gaps.



As you can see, when I use groundingdino + sam2 to get the annotation and mask of the portrait image. I use the prompt of "person. headphones". The box result of Dino is correct, but in the segmentation result of Sam2, there is a gap between the headphones and the human body, and there is an headphone line in the human body that is not segmented. Such a mask cannot be used for portrait segmentation tasks. How can I deal with it? Can it be solved by post-processing or do I need to finetune Sam2?