Image segmentation is the process of separating an image into object and nonobject components. Computer vision applications like medical and biomedical imaging, satellite imaging, robot navigation, and image retrieval are based on the precise segmentation of images.
The authors present a semantic image segmentation method using structural support vector machines (S-SVMs) to label image components. They introduce inter-class co-occurrence statistics to maintain the multi-label ranking consistency. The method uses the multi-label ranking score, making it a global constraint distinguishing between similar and dissimilar component labels.
The method is made simple by decoupling global with pairwise co-occurrence information and combining it with local unary features and logistic regression. The authors used the Microsoft Research Cambridge-21 (MSRC-21) dataset and the Stanford background dataset (SBD) for the experiments. With fivefold cross-validation performing over the randomly divided SBD, the authors claim that the method is highly competitive compared to state-of-the-art methods.
Semantic analysis of the component regions in an image-discretizing multi-label ranking makes the paper worth reading. The authors present the image semantics concept in a very informative manner and would like to add depth as an extra feature into S-SVM frameworks for future improvements.