Image aesthetics assessment (IAA) has been an important topic in computer vision research. In recent years, many deep learning based approaches have been developed for machine automatic image (photograph) assessment. In this paper, we study the IAA for photography assessment, which captures the most significant aesthetic features in photography. In particular, we re-examine the multi-column deep convolutional neural network architecture, in which photographs are assessed based on their global and local views to make the assessment more precise. Our experiments are conducted with a completely new dataset we have built - Curated Photography Dataset (CPD) which contains over 500,000 photographs crossing eight different categories, and all of these photos have been curated by professional photographers and curators. We show that our approach outperforms the state of the art approaches in the area, and sheds a new light for developing practical AI photography curators in real world domains.