A Self-supervised Pretraining Method Based on Mixup for Ordinal Classification

Ordinal classification is an important research topic that assigns instance with ordinal category. In practical applications, it greatly relies on supervised deep models, which require massive amount of labeled samples. In this paper, for small or moderate dataset, we use self-supervised learning(SSL) as a pretraining process to guide supervised learning. In self-supervised pertraining, rotating image embedded with Mixup strategy enables model to learn richer feature representation. Pretrained model is learned by rotation difference between two combined images, instead of single image rotation. After pretraining, SSL helps supervised learning with the style of finetuning. Finally, we evaluate the effectiveness of pretraining paradigm and Mixup rotation strategy on two datasets (Adience and Carstyle) in the experiment, and achieve promising gains on classification performance.