Learning Complex Action Patterns with

摘要

This paper deals with the problem of automatically compiling rules which describe complex actions in terms of the spatio-temporal attributes of labeled parts. Of particular interest is the exploration of a model-based approach to induction of part attributes constrained by known properties of the generation process. The resultant algorithm is based on constraint propagation over spatio-temporal decision trees which produces Horn clause descriptions which depict the spatio-temporal properties of parts and their relations which satisfy training conditions.