how to obtain classification accuracy as mentioned in the paper? per pixel accuracy and class accuracy.