Infinite Latent Conditional Random Fields

摘要

In this paper, we present Infinite Latent Conditional Random Fields (ILCRFs) that model the data through a mixture of CRFs generated from Dirichlet processes. Each CRF represents one possible explanation of the data. In addition to visible nodes and edges that exist in classic CRFs, it generatively models the distribution of different CRF structures over the latent nodes and corresponding edges, imposing no restriction on the number of both nodes and types of edges. We apply ILCRFs to several applications, such as robotic scene arrangement and scene labeling, where a scene is modeled through, not only objects, but also latent human poses and human-object relations. In extensive experiments, we show that our model outperforms the state-of-the-art results as well as helps a robot placing objects in a new scene.