Statistical Adaptive Metric Learning for Action Feature Set Recognition in the Wild

This paper proposes a statistical adaptive metric learning method by exploring various selections and combinations of multiple statistics in a unified metric learning framework. Most statistics have certain advantages in specific controlled environments, and systematic selections and combinations can adapt them to more realistic “in the wild” scenarios. In the proposed method, multiple statistics, include means, covariance matrices and Gaussian distributions, are explicitly mapped or generated in the Riemannian manifolds. Subsequently, by embedding the heterogeneous manifolds in their tangent Hilbert space, the deviation of principle elements is analyzed. Hilbert subspaces with minimal principle elements deviation are then selected from multiple statistical manifolds. After that, Mahalanobis metrics are introduced to map the selected subspaces back into the Euclidean space. A uniformed optimization framework is finally performed based on the Euclidean distances. Such a framework enables us to explore different metric combinations. Therefore our final learning becomes more representative and effective than exhaustively learning from all the hybrid metrics. Experiments in both static and dynamic scenarios show that the proposed method performs effectively in the wild scenarios.