UC Merced

We propose two methods for object segmentation by combining learned shape priors with local features. The first, Max-Margin Boltzmann Machines, learns shapes in an unsupervised way, followed by a joint refinement using features extracted from the image, using max-margin methods. Second, we investigate the feasibility of another approach, based on deep learning and patchwise output mask refinement. As a way to further improve results, we also present an application of structured learning to learn Graph Cut based segmentation mask smoothing.

We conduct experiments on datasets containing diverse images of three classes of objects, showing promising results. We also discuss both qualitative and quantitative results extensively and point out both the strengths and shortcomings of the above approaches.