Focusing on your subject: Deep subject-aware image composition recommendation networks

Abstract Photo composition is one of the most important factors in aesthetics photographs. As a popular application, recommendation for photo focusing on specific subject has been ignored by recent deep-learning-based approaches. In this paper, we propose subject-aware image method, SAC-Net, which takes an RGB and binary window mask as input, returns good compositions crops containing subject. Our model first determines candidate scores all possible coarse cropping windows. The with high are selected further refined regressing their corner points to generate output recommended final predicted score regression module. Unlike existing methods that need preset several windows, our network able automatically regress windows arbitrary aspect ratios sizes. We novel stability losses maximizing smoothness when changing along view changes. Experimental results show method outperforms state-of-the-art not only task, but also general purpose recommendation. have designed multistage labeling scheme so large amount ranked pairs can be produced economically. use dataset SACD, contains 2777 images, more than 5 million pairs. SACD publicly available at https://cg.cs.tsinghua.edu.cn/SACD/ .