Find Beauty in the Rare: Contrastive Composition Feature Clustering for Nontrivial Cropping Box Regression

Automatic image cropping algorithms aim to recompose images like human-being photographers by generating the boxes with improved composition quality. Cropping box regression approaches learn beauty of from annotated boxes. However, bias annotations leads quasi-trivial recomposing results, which has an obvious tendency average location training samples. The crux this predicament is that task naively treated as a problem, where rare samples might be dominated normal samples, and patterns are not well exploited. Observing similar tend shared boundaries nearly, we argue find clustering boundary annotations, i.e., patterns. We propose novel Contrastive Composition Clustering (C2C) regularize features contrasting dynamically established dissimilar pairs. In way, common multiple can better summarized, especially benefits endows our model generalizability render nontrivial results. Extensive experimental results show superiority compared prior arts. also illustrate philosophy design interesting analytical visualization.