[Query] about crop images in CLIP prompts

[gasharper]

In the text_promt function, the first return value cropped_boxes is used in post retrieve function. However, the first return value comes from _segment_image, which leaves a large white padding for most small bboxs.

Why not crop the image by the bounding box, then put them into CLIP? Does white padding affect CLIP results?

[AtiChetsurakul]

I just getting take a look in the cropped_boxes and wonder if it some solution, bug or essential for something. I think It might be better if we change their function _crop_image to just use PIL.Image.crop ?