Competitive-Driven Learning for Image Ordinal Classification

Image ordinal classification (IOC) assigns a discrete yet ordinal scalar label to an input image, such as age estimation. For IOC, it is non-trivial to effectively incorporate the ordinal information of inputs into the classification task. In this work, we adopt the concept of CutMix, a data augmentation technique that mixes two samples and their labels proportionally during the training, to help the task of ordinal classification. In particular, we study the CutMix from the perspective of metric learning instead of data augmentation, which naturally and simply implements classification learning and difference regression learning. Firstly, we embed two ordinal images into the training procedure by a virtual combination of image and label, and explain why CutMix fits IOC well. Secondly, the ordinal difference between two images is utilized to learn the fine difference in category classification, which is integrated with CutMix properly. Finally, for further improvement, dual pairing and random augmentation are proposed to enlarge the disparity of two objects, so as to exploit more discriminative feature. Comprehensive experiments demonstrate the effectiveness of our newly proposed approach, which outperforms previous methods with a relatively large margin (>2%).