I’m doing a binary segmentation task.In the image pre-processing procedure, the image and mask are resized to same size.But the problem is: mask is a binary value image,its pixels represent a label ,so it should resized with NEAREST so the label value is still 0 or 1. If I use bilinear or cubic interpolation method, the resized mask image
contains some decimal number ,and after a ‘mask.long()’ ops,these decimal all becoms 0, it do affect the mask which are sented into the network for loss calculation,especially on some small size object, their ground truth image turn out like a scatter map.
Should I use nearest to interolate the image pair? is there any other solution besides nearest?
the following image is a triple image example,from left to right is original image ,segmentation result(my result),ground truth,
actually, the ground truth can’t be worse more.
