Multi-model fusion metric learning for image set classification

摘要

Multi-model image set classification is an important research topic in pattern recognition, because multiple modeling of image sets can represent different characteristics. This paper focuses on the problem of two-model image set classification. We use the mean vector, subspace and covariance matrix to jointly represent an image set, due to they contain different discriminative information. Since the above three representation methods lie on different spaces, the classical multi-view learning methods cannot be directly utilized. In order to reduce the dissimilarity between the heterogeneous spaces, a new multi-model fusion metric learning (MMFML) framework is developed, which includes three two-model fusion Modes. Specifically, by exploiting corresponding kernel function, the original represent data is first embedded into the high dimensional Hilbert spaces to reduce the gaps between the heterogeneous spaces. The final objective function is then learned from the Hilbert space to a common Euclidean subspace, and in the final subspace the classical Euclidean distance is used to classify image sets. Experimental results on Honda/UCSD, Youtube Celebrities and ETH-80 three databases demonstrate the effectiveness of our proposed method.