Kernelised reference‐wise metric learning

Abstract

Unlike the doublet or triplet constraints, a novel kernelised reference-wise metric learning is proposed by constructing reference-wise constraints, which contain similarity information of each sample to all reference samples. After selecting several representative training samples as the reference set, the training data are first mapped into a projected space by the reference sample matrix. Then the problem of metric learning is casted as a multi-label classification problem under the reference-wise constraints, in which an l 2-norm regularised least squares canonical correlation analysis can be used. Besides, the formulation is generalised to kernelised version for further boosting the performance. Experiments on two benchmark person re-identification datasets demonstrate that the approach clearly outperforms the state-of-the-art methods.