The fine-tuned HQ-SAM model shows a significant improvement in accuracy. However, the val_iou_0 accuracy of HQ-SAM during training is very low. about sam-hq HOT 2 CLOSED

Hi, if the visualization result is correct but the IoU metric is strange, this may be because the format of GT or prediction during testing is incorrect.
You can check the input when testing the metric. For example, we compute IoU this line. masks_hq has a range of (-inf,+inf), labels_ori has a range of [0,255]. You can check the range of some variables, or visualize some intermediate variables to determine whether there is a problem with the format.

from sam-hq.

@ymq2017 , Thanks.

I found that the key is in the line, which states that the input image dimensions must be [1024,1024]. If I modify the "input_size" to any other value, the IoU will be incorrect.

from sam-hq.