Abstract:

We present algorithms for automatic feature selection and model order identification based on our previous solution to unsupervised structure discovery from video sequences. The overall problem is presented as simultaneously finding the statistical descriptions of structure and locating segments that matches the descriptions. Structures in video was modelled with hierarchical hidden Markov models, and model parameters was efficiently learned using EM. We extend the previous model adaptation scheme to learning not only the complexity of each structure, but also the total number of structures in the stream. Feature selection iterates between a wrapper and a filter method to partition the large feature pool into consistent and compact subsets, where the subsets are then ranked according to a normalized Bayesian Information criteria. Results on soccer videos are very promising: the best feature set agrees with manually identified significant features, the clusters are explainable with respect to manual labels, and the accuracy is comparable with previous works with supervised learning or manually chosen feature sets.