An MDL Approach to Learning Activity Grammars

Abstract

Stochastic Context-Free Grammars (SCFG) have been shown to be useful for vision-based human ac-tivity analysis. However, action strings from vision-based systems differ from word strings, in that a string of symbols produced by video contains noise symbols, making grammar learning very difficult. In order to learn the basic structure of human activities, it is necessary to filter out these noise symbols. In our work, we propose a new technique for identifying the best subset of non-noise terminal symbols and acquiring the best activity grammar. Our approach uses the Minimum Description Length (MDL) principle, to evaluate the trade-offs between model complexity and data fit, to quantify the difference between the results of each terminal subset. The evaluation results are then used to identify a class of candidate terminal subsets and grammars that remove the noise and enable the discovery of the basic structure of an activity. In this paper, we present the validity of our proposed method based on experiments with synthetic data.