Fig. 2From: Pedestrian segmentation based on a spatio-temporally consistent graph-cut with optimal transportFramework of the ESS. a Each pixel in the input image (e.g., a 5×5 grayscale image) initialized as a superpixel, where a black number is the label of a superpixel. b Each pixel relabeled under an energy minimization framework. In each iteration, we scan and update the labels of all pixels. For each pixel (yellow), the label assignment costs of its four-connected neighbors (blue) are calculated as shown by red numbers, and each pixel’s label is updated with the lowest-cost neighbor’s label. The iteration continues until there is no change in each pixel’s label. Finally, the superpixel segmentation result is obtained as in cBack to article page