Exploiting Richness of Learned Compressed Representation of Images for Semantic Segmentation