Studies On Triplet Distance Constrained Metric Learning

Distance metric learning is an important research topic in machine learning and computer vision. It aims to learn a proper distance metric, which makes the distance between samples from the same class smaller and the distance between samples from different classes larger. The performances of many machine learning algorithms depend on their distance metrics. Distance metric learning has been applied to many real applications, e.g. classification, clustering, retrieval and identification.Traditional metric learning methods represent samples as points in high dimensional space and impose doublet or triplet constraints on the se points. However, in many circumstances, the data we obtained can be obtained into sample sets, e.g. video image sets, etc. Based on this, we develop a metric learning model based on triplet distance constraints, and we extend it to point-to-set and set-to-set distance metric learning. Our work include the following aspects:We propose a point-to-point relative distance constrained metric learning(RDCML) model. For each triplet(xi,xj,xk),RDCML restricts the distance between samples xi and xj with the same class label should be smaller than the distance between xi and xk with different class labels. We also develop an SVM-like algorithm to solve the RDCML model. Experimental results demonstrate that RDCML can achieve better or comparable performance competing with the state-of-the-art metric learning algorithms.We extend RDCML to point-to-set distance metric learning. By computing the point-to-set distance, we construct the triplets and learn the metric matrix M due to the triplet constraints. We compute the minimum distance between a sample point and the subspace that a set spanned, and use the distance to do classification. Experiment results on object recognition and face recognition prove the efficiency of the method.We also extend RDCML to set-to-set distance metric learning. We construct triplets by computing the set to set distance, then iterate between triplet construction and distance metric learning to solve the problem. Experiment results on object recognition and face recognition dataset demonstrate that our method can achieve leading performance. We also collect a small tongue image set database, result on this database shows our method is capable for tongue classification, which is a good start for future research.