Algorithms for inference in Specialized Maps for recovering 3D hand pose

Learning in the Specialized Mappings Architecture (SMA) was presented for recovering 3D hand pose from visual features, however inference was not fully addressed. We tackle this aspect of the SMA model more thoroughly, and propose two inference algorithms, one deterministic and one probabilistic. SMA consists of several specialized forward (input to output) mapping functions that are estimated automatically from data and a known feedback or inverse function (that maps outputs to inputs). This allows the use of alternative conditional independence assumptions in the same model (derived from a forward and a feedback model) during learning and inference. The two proposed inference approximations, the mean output (MO) and the multiple sampling (MS) algorithms, are tested in recovering 3D hand pose from single images