Large scale person re-identification : error rates minimization under time constraint

Details

Supervisors

Dupont, Pierre

Faculty

Ecole polytechnique de Louvain

Degree label

Master [120] : ingénieur civil en informatique

Abstract

Given a population of individuals, the task of person re-identification consists of identifying the person inside the population to which corresponds the test data, if this person exists. In a large scale setting, realistic time and space constraints have to be posed, both for the learning and decision processes. In this thesis, the person re-identification problem is solved by a threshold nearest neighbour search. The decision threshold minimizing the error rates has been shown to be dependent on the population size. With such an adaptive decision rule, the error rates increase sublinearily with the number of individuals, which allows to tackle large populations. In order to meet the time constraint, the decision rule has to be adapted, which degrades the identification performance. It is thus crucial to develop search algorithms, but also metrics allowing for an efficient search. A Mahalanobis metric learning scheme balancing between identification performance and computational efficiency will be proposed. The compromise is done by casting a convex optimization problem, that can be solved using a simple gradient descent algorithm. For some parameter values, this method is showed to be equivalent to standard metric learning for person re-identification. Our method has shown a drastic improvement of the search efficiency on partially artificial data, compared to standard margin maximization metrics. In presence of a hard time and space constraint, this leads to a significant enhancement of the identification performance.