Multi-view Viewpoint Assessment for Architectural Photos.

This paper proposes a robust method of viewpoint assessment for taking a good photograph of architecture. Unlike the conventional works devoted to assessing the aesthetic of a photograph mainly relying on the image features, both the image and geometric features are extracted from the architecture photos in our method. Furthermore, we explore the mutual knowledge between these two aspects of features with multi-view learning. With the learner trained by multi-view learning, the viewpoint goodness of architecture photograph can be assessed by either aspect of the features. Experiments suggest that the multi-view learning with kernel canonical correlation analysis achieves superior performance over using solely traditional image features. With the help of multi-view learning, we can harness the geometric cues with image features effectively.