UC Santa Cruz

The automatic evaluation of images in computational cinema and photography is a challenging problem. There are neither comprehensive rules nor adequate training data to develop an expert system. This thesis presents the design and implementation of a computer game for synthesizing image data, and an experimental framework for learning photographic composition preferences from online ratings.

The first topic addressed is the development of Panorama, a computer game for photographic composition research. Panorama generates images through gameplay by simulating the task of photography in a virtual environment. The synthesized photographs are automatically annotated from their underlying representation, and contribute to a corpus of images with well-defined visual features. Both the game and data are publicly available to support research in image analysis.

The second topic is a photograph learning experiment using data from the Panorama corpus. In this section, machine learning is used for predicting user preferences of images, based on computed visual features. Image preference ratings are acquired by crowdsourcing to consider subjectivity across many individuals. Using this unique data collection process, it is shown how the framework can be applied to reason about image quality with respect to photographic composition.