ROI for input images
Kif11 opened this issue · comments
It would be useful to specify ROI (Region of interest) for a given image that the image get pre-cropped before feeding into the net. However the final output resolution stay the same as input image.
This is useful when foreground object is small compare to background. Cropping the image helps with mask quality like on image bellow:
This is a good idea, adding it to feature requests.