Distilling Knowledge From Object Classification to Aesthetics Assessment

In this work, we point out that the major dilemma of image aesthetics assessment (IAA) comes from abstract nature aesthetic labels. That is, a vast variety distinct contents can correspond to same label. On one hand, during inference, IAA model is required relate various other when training, it would be hard for learn distinguish different merely with supervision labels, since labels are not directly related any specific content. To deal dilemma, propose distill knowledge on semantic patterns multiple pre-trained object classification (POC) models an model. Expecting combination POC provide sufficient contents, easier limited number By supervising end-to-end single-backbone distilled knowledge, performance significantly improved by 4.8% in SRCC compared version trained only ground-truth categories images, improvement brought proposed method achieve up 7.2%. Peer comparison also shows our outperforms 10 previous methods.