Structured Completion Predictors Applied to Image Segmentation

Multi-image segmentation makes use of global and local features in an attempt to classify every pixel in an image into a semantic region. One important relationship is inter-class spacial interaction between small local regions we call “superpixels”. While complex models have been built in order to provide for such semantic understanding and define visual grammars, we explore the usefulness of a new technique we name stacking. This method performs multi-class SVM learning several times by first constructing label probabilities for each superpixel based on extremely local superpixel features, such as raw pixel information, and increases semantic understanding through stacked predictors that augment the feature set with predictions on the surrounding context. The results of the algorithm with our constructed features are marginally successful, demonstrating that the algorithm can be used to improve semantic understanding. This report analyzes the success of constructed features and analyzes reasons for a lack of strong improvement.