Quantitative Analysis of Automatic Image Cropping Algorithms: A Dataset and Comparative Study

摘要

Automatic photo cropping is an important tool for improving visual quality ofdigital photos without resorting to tedious manual selection. Traditionally,photo cropping is accomplished by determining the best proposal window throughvisual quality assessment or saliency detection. In essence, the performance ofan image cropper highly depends on the ability to correctly rank a number ofvisually similar proposal windows. Despite the ranking nature of automaticphoto cropping, little attention has been paid to learning-to-rank algorithmsin tackling such a problem. In this work, we conduct an extensive study ontraditional approaches as well as ranking-based croppers trained on variousimage features. In addition, a new dataset consisting of high quality croppingand pairwise ranking annotations is presented to evaluate the performance ofvarious baselines. The experimental results on the new dataset provide usefulinsights into the design of better photo cropping algorithms.