Effective Aesthetics Prediction with Multi-level Spatially Pooled Features

We propose an effective deep learning approach to aesthetics qualityassessment that relies on a new type of pre-trained features, and apply it tothe AVA data set, the currently largest aesthetics database. While previousapproaches miss some of the information in the original images, due to takingsmall crops, down-scaling or warping the originals during training, we proposethe first method that efficiently supports full resolution images as an input,and can be trained on variable input sizes. This allows us to significantlyimprove upon the state of the art, increasing the Spearman rank-ordercorrelation coefficient (SRCC) of ground-truth mean opinion scores (MOS) fromthe existing best reported of 0.612 to 0.756. To achieve this performance, weextract multi-level spatially pooled (MLSP) features from all convolutionalblocks of a pre-trained InceptionResNet-v2 network, and train a custom shallowConvolutional Neural Network (CNN) architecture on these new features.