Search

Deep convolutional networks are data hungry learners and, to compensate for the limited amount of available data, various augmentation methods have been proposed. While the initial approaches aimed to fill the space around existing data points, more recent methods use the "mixing" procedure, which is a convex combination between points, being both holistic and local (from a spatial point of view). Although the mixing techniques improved the performance for standard benchmarks, we do notice that they are less effective for problems that exhibit poor class separation. For these scenarios, we propose a soft labeling technique based on distances to class prototypes, as extracted on an intermediate CNN layer. The mixing is performed between relative positions in the clustering space. The method, named SoftClusterMix, is shown to be competitive on standard image classification benchmarks and leads to significantly improved accuracy for problems where there is a poor class separation. We report better performance in three such categories: face expression recognition, aesthetic image classification and painting style classification. Ablation tests allow deep insights of the method. [ABSTRACT FROM AUTHOR]