Learning Computationally Efficient Metrics for Large Scale Person Identification
(2018) BENELEARN2018 — Location: Jheronimus Academy of Data Science (8.November.2018)
Files
Benelearn_2018_paper_4.pdf
Open Access - Adobe PDF
- 822.91 KB
Details
- Authors
- Abstract
- Nearest neighbor (NN) classifiers rely on a distance metric either a priori fixed or previously estimated through metric learning. When such metric learning occurs, a natural objective is to minimize the classification errors of the NN classifier. This learning procedure is however commonly performed without any regard to the computational efficiency of the NN classifier at test time. In this work, we propose to formulate the metric learning problem as a multi-objective trade-off between classification performance and computational efficiency at test time. This is illustrated here in the context of person identification. Specifically, a Mahalanobis metric learning scheme is cast as a convex optimization problem over a set of positive semi-definite matrices, and solved through projected gradient descent. Experimental results are presented on semi-artificial data, representative of a profile-based person identification task at a large scale (> 10^6 individuals). Our method shows a significant improvement of the search efficiency of a NN classifier, compared to standard soft-margin maximization metrics. In presence of hard time and space constraints, it leads to a drastic enhancement of the identification performance.
- Affiliations
Citations
Hamer, V., & Dupont, P. (2018). Learning Computationally Efficient Metrics for Large Scale Person Identification. Proceedings of the Annual Machine Learning Conference of Belgium and the Netherlands 2018. Published. BENELEARN2018, Jheronimus Academy of Data Science. https://hdl.handle.net/2078.5/253691 (Original work published 2018)