Boosted Metric Learning for Efficient Identity-Based Face Retrieval

International audienceThis paper presents MLBoost, an efficient method for learning to compare face signatures , and shows its application to the hierarchical organization of large face databases. More precisely, the proposed metric learning (ML) algorithm is based on boosting so that the metric is learned iteratively by combining several weak metrics. Boosting allows our method to be free of any hyper-parameters (no cross-validation required) and to be robust with respect to overfitting. This MLBoost algorithm can be trained from constraints involving two pairs of vectors (quadruplets) with a quadratic complexity. The paper also shows how it can be included in a semi-supervised hierarchical clustering framework adapted to identity based face search. Our approach is validated on a benchmark relying on the Labelled Faces in the Wild (LFW) dataset supplemented with 1M face distractors