When i submit the results I obtain a really low IoU (0.01) but during training (on validation set) It Is around 60-70%. I think the problem is that i resized the images to 256×256 but for the evaluatio It Is required a full size image prediction. I also tried to resize directly the prediction from 256×256 to full size without interpolation but the result was still bad.
Does anyone have the same problem?