Deep Segmentation Networks Using "Simple" Multi-Layered Graphical Models

This paper provides a new perspective on a number of traditional unsupervised segmentation methods, and a number of more recent segmentation methods, in terms of multi-layered graphical models and supervised learning. This perspective suggests a family of segmentation methods, which we call deep segmentation networks that are different, but complementary, to the typical deep network used for classification. One of the biggest differences in DSNs is the pooling layer. Deep classification networks typically use pre-specified fixed pooling regions, while deep segmentation networks use data adaptive pooling regions. We investigate some of the architectural choices with these new architectures in experiments with benchmark data and suggest directions for future work.