Hi,
I use Mask R-CNN to detect and segment large objects (about 750x750 pixels) in a very large image. The detection works really well, but the segmentation seems a little bit approximative on the borders.
Mask R-CNN has a segmentation branch that segment a m x m mask for each ROI. This m x m mask is usually small (about 28x28 if I’m right) and after passing through the ROI_Align it has to represent/segment a big 750x750 ROI.
Is there a way to modify the dimensions of the small m x m mask to make it bigger?