Learning Depth in Light Field Images

Photographic images reduce the three dimensional world they are capturing into a 2D plane. There are a variety of applications in which it would be useful to determine the original distance to objects in the scene from the focal plane of the image. Such applications include robotic vision systems, 3D scene reconstruction, and surveying. Recent efforts to provide automatic analysis of scene depth using a single image have proved quite successful[4], however, ambiguities still arise in complex scenes, and non-general assumptions of the environment are made in ensure the best response of the algorithms. In contrast to traditional cameras, recent work has made use of a light field camera, which captures the 4D light field of the scene. The details of how this is done are described below. By using the extra information in the light field, we hope to provide a more general implementation of a depth map learning algorithm, which requires fewer parameters to train, and which will work in a variety of environments.