MLBoost Revisited: A Faster Metric Learning Algorithm for Identity-Based Face Retrieval

International audienceThis paper addresses the question of metric learning, i.e. the learning of a dissimilar-ity function from a set of similar/dissimilar example pairs. This domain plays an important role in many machine learning applications such as those related to face recognition or face retrieval. More specifically, this paper builds on the recent MLBoost method proposed by Negrel et al. [25]. MLBoost has been shown to perform very well for face retrieval tasks, but this algorithm relies on the computation of a weak metric which is very time consuming. This paper demonstrates how, by introducing sparsity into the weak projectors, the convergence time can be reduced up to a factor of 10× compared to MLBoost, without any performance loss. The paper also introduces an explicit way to control the rank of the so-obtained metrics, allowing to fix in advance the dimension of the (projected) feature space. The proposed ideas are experimentally validated on a face retrieval task with three different signatures