An adaptive ensemble-based system for face recognition in person re-identification

摘要

Recognizing individuals of interest from faces captured with video cameras raises several challenges linked to changes in capture conditions (e.g., variation in illumination and pose). Moreover, in person re-identification applications, the facial models needed for matching are typically designed a priori, with a limited amount of reference samples captured under constrained temporal and spatial conditions. Tracking can, however, be used to regroup the system responses linked to a facial trajectory (facial captures from a person) for robust spatio-temporal recognition, and to update facial models over time using operational data. In this paper, an adaptive ensemble-based system is proposed for spatio-temporal face recognition (FR). Given a diverse set of facial captures in a trajectory of a target individual, an ensemble of 2-class classifiers is designed. A pool of ARTMAP classifiers is generated using a dynamic PSO-based learning strategy, and classifiers are selected and combined using Boolean combination. To train classifiers, target samples are combined with a set of reference non-target samples selected from the cohort and universal models using One-Sided Selection. During operations, facial trajectories are captured, and each individual-specific ensemble of the system seeks to detect target individuals, and possibly self-update their facial models. To update an ensemble, a learn-and-combine strategy is employed to avoid knowledge corruption, and a memory management strategy based on Kullback-Leibler divergence allows to rank and select stored validation samples over time to bound the system's memory consumption. Spatio-temporal fusion is performed by accumulating classifier predictions over a time window, and a second threshold allows to self-update facial models. The proposed systems were validated with videos from the Face in Action and COX-S2V datasets, that feature both abrupt and gradual patterns of change. At the transaction level, results show that the proposed system allows to increase AUC accuracy by about 3 % for scenarios with abrupt changes, and by about 5% with gradual changes. Subject-based analysis reveals the difficulties of face recognition with different poses, affecting more significantly the lamb- and goat-like individuals. Compared to reference spatio-temporal fusion approaches, results show that the proposed accumulation scheme produces the highest discrimination.