The algorithm described in the paper "Deep Neural Network ensembles dedicated to different head poses for face identification" was proposed to resolve the task of the of the ICB-RW 2016 challenge (International Challenge on Biometric Recognition in the Wild).

Probe images from ICBRW dataset.

The main objective of the competition was to identify the people in the CCTV surveillance system assuming that the watch-list with good quality images was provided. Training dataset, which was delivered by the organizers, contained both the watch-list images of 90 subjects (called there the gallery subset) and the probe images which are the static frames obtained from the surveillance system. Each subject in in the gallery subset was represented by the 3 images with different poses (frontal, left-side and right-side) The probe images was captured outdoor, subjects rarely look into the camera that was localized over their heads and faces are often occluded.
The results of ICB-RW Challenge were evaluated using an identification index given by the area under curve (AUC) calculated for the cumulative match score (CMC). The proposed algorithm took fourth place in the challenge.

Challenge evaluation results.

Method was based on deep neural networks ensembles. Each ensemble network was trained to classify faces seen from the specific directions, namely, frontal images, left and right profiles. This approach meant to reflect the given watch list structure. Results were then combined to obtain the final classifier.

Detailed description in papers:

1. J Neves, H Proença, "ICB-RW 2016: International challenge on biometric recognition in the wild", Biometrics (ICB), 2016 International Conference on, 1-6
2. H. Proença, M. Nixon, M. Nappi, E. Ghaleb, G. Özbulak, H. Gao, H. K. Ekenel, K. Grm, V. Struc, H. Shi, X. Zhu, S. Liao, Z. Lei, S. Z. Li, W. Gutfeter, A. Pacut, J. Brogan, W. J. Scheirer, E. Gonzalez-Sosa, R. Vera-Rodriguez, J. Fierrez, J. Ortega-Garcia, D. Riccio, L. De Maio, "Trends and Controversies", IEEE Intelligent Systems, vol. 33, no. 3, pp. 41-67, May./Jun. 2018.