Image Understanding in a Nonparametric Bayesian Framework

We explore recently proposed nonparametric Bayesian statistical models of image partitions. These models are attractive because they adapt to images of different complexity, successfully modeling uncertainty in size, shape, and structure of human segmentations of natural scenes. We improve upon them in a number of key ways to achieve performance comparable to state-of-the-art methods. Our first major contribution is a novel discrete search based posterior inference algorithm which, compared to previous approaches, is significantly more robust and accurate. We then present a low rank version of the spatially dependent Pitman-Yor processes model, critical for efficient inference. Furthermore, we show how the Gaussian process covariance functions underlying the proposed models can be calibrated to accurately match the statistics of human segmentations. Finally, we present accurate segmentations of complex scenes as well as multiple hypothesized image partitions (capturing the inherent uncertainty in human scene interpretations) produced by our method.