Testing Set The testing set was composed by 52 images. The distance metric was learned from the training set and applied here to segment these images. At left we present the original images used in our experiments. At center we show the segmentation results of the Mumford-Shah algorithm when using the simple Euclidean distance to compute the similarity between the color vectors. At right, the segmentation results of the proposed approach using the training set. This testing set has zero overlap over the training set. Original Image Euclidean Distance Proposed Approach