Vistas: Hierarchial Boundary priors using Multiscale Conditional Random Fields

Detection of natural boundaries is a fundamental problem in computer vision but evaluation of boundary detection performance has tended to concentrate on images with low scene complexity. Importantly, recent boundary detection analysis [7] shows that performance on scenes with higher scene complexity is low. However, work in [6] has shown that for datasets with some scene consistency [1] it is possible to learn a distribution over the density of boundaries in the image. The work exploits the fact that when a 2D image is a projection of a 3D scene, perspective effects result in an uneven distribution of the sizes of object classes, and therefore an uneven distribution in the density of object boundaries across the scene. This observation is illustrated in Figure 1a-b. This suggests learning a non-stationary model for boundary priors. In the paper we achieve this by combining two methods that have proved effective for image labeling problems [3, 4] and learn a mixture of multiscale conditional random fields. The basis of our CRF is a generative model over observed boundaries x = [x1 . . .xP] at the p pixels of an image patch. xp is binary and is 1 when a boundary is present and 0 when it is absent. We use a clustered sub-space model introduced in [6] which models the probability of x with an activation a = μc + Fch comprising a mean, μ , factor matrix, F, and hidden variable h. This model is learned using the EM algorithm. We can summarize the model concisely (see Figure 1c) as: