Dictionary pair learning in compressed space for action recognition

Action recognition is still a challenging problem. In order to catch effective compact representation of the action sequences, the discriminative dictionaries could be learned by sparse coding. But sparse coding is needed in both the training and testing phases of the classifier framework. And it is also time consuming for the adoption of 1-norm sparsity constraint on the representation coefficients in most dictionary learning (DL) methods. Dictionary pair learning (DPL) learns a synthesis dictionary and an analysis dictionary jointly. Compared with those DL approaches, the using of DPL method may not only effectively reduce the time consuming during the phases of training and testing, but also result in very competitive recognition ratio. On the other hand, the way of compressed learning can lead to learning with randomly projected data instead of original data. Thus compressed learning could greatly cut down on both the requirement of memory storage and running time due to the effective reduction of data dimensions through random projection. In this paper, Combined with compressed learning, DPL in compressed space are explored for the action recognition. By the experiments on various public action datasets, it has been shown that DPL in compressed space can achieves very competitive accuracy, while it is significantly faster in phases of both training and testing, which indicate the efficiency of the proposed algorithm for action recognition.