Adaptive similarity measures for matrix objects based on feature variation and sequence length for gesture recognition

In pattern recognition, the usage of appropriate similarity measure is crucial for acquiring robust performance. In this paper, we present two similarity measures, which compute a similarity between two matrices, based on the temporal feature variation and sequence length difference, respectively. Especially, the matrix object used in this paper has unique characteristic. Each row and column has different meaning. Each row corresponds to a frame of sequence and each column corresponds to variation of a feature by time. We analyze this characteristic of matrix object and extract additional information for acquiring robust performance. From these analyses, we assume that the feature with a large amount of variation is more important than the feature with lower variation and the sequence length difference between two gestures is helpful to calculate a similarity. And then we define two factors: feature importance factor and scaling factor. And then we present two similarity measures and apply them to three public benchmark databases: Cambridge hand gesture database, ChaLearn database and SKIG database. The experimental results show that our proposed similarity measures acquires the improvement of performance.