Rule Induction for Adaptive Sport Video Characterization Using MLN Clause Templates

The grounding of high-level semantic concepts is a key requirement of video annotation systems. Rule induction can thus constitute an invaluable intermediate step in characterizing protocol-governed domains, such as broadcast sports footage. We here set out a novel “clause grammar template” approach to the problem of rule-induction in video footage of court games that employs a second-order meta-grammar for Markov Logic Network construction. The aim is to build an adaptive system for sports video annotation capable, in principle, both of learning ab initio and also adaptively transferring learning between distinct rule domains. The method is tested with respect to both a simulated game predicate generator and also real data derived from tennis footage via computer-vision based approaches including HOG3D based player-action classification, Hough-transform-based court detection, and graph-theoretic ball-tracking. Experiments demonstrate that the method exhibits both error resilience and learning transfer in the court domain context. Moreover the clause template approach naturally generalizes to any suitably-constrained, protocol-governed video domain characterized by feature noise or detector error. keywords: Video Annotation, Markov processes, Stochastic Logic, Markov logic network (MLN), Action Recognition, Behavior discovery, Statistical Relational Reasoning