← Convolutional Random Walk Networks for Semantic Image Segmentation Puzzle-CAM Improved localization via matching partial and full features→