Exploiting representation plurality for robust and efficient face recognition Vitomir ˇ Struc 1 , Jerneja ˇ Zganec Gros 2 , Simon Dobriˇ sek 1 , Nikola Paveˇ si´ c 1 1 Faculty of Electrical Engineering, University of Ljubljana Trˇ zaˇ ska 25, SI-1000 Ljubljana, Slovenia 2 Alpineon d.o.o., Ulica Iga Grudna 15, SI-1000 Ljubjana, Slovenia E-mail: {vitomir.struc, simon.dobrisek, nikola.pavesic}@fe.uni-lj.si, jerneja.gros@alpineon.si Abstract The paper introduces a novel approach to face recogni- tion that exploits plurality of representation to achieve robust face recognition. The proposed approach was sub- mitted as a representative of the University of Ljubljana and Alpineon d.o.o. to the 2013 face recognition compe- tition that was held in conjunction with the IAPR Inter- national Conference on Biometrics and achieved the best overall recognition results among all competition partic- ipants. Here, we describe the basic characteristics of the submitted approach, elaborate on the results of the com- petition and, most importantly, present some general find- ings made during our development work that are of rele- vance to the broader (face recognition) research commu- nity. 1 Introduction Despite the progress made during the last decades, the ex- isting face recognition technology still suffers from some drawbacks that relate mainly to the use of the technology in uncontrolled and unconstrained environments. In such environments the technology commonly achieves much lower recognition rates than under controlled and super- vised conditions. Two dominant research directions have emerged to address this problem. The first focuses on extracting in- variant yet discriminative facial features which can be used with robust classification approaches to cope with facial appearance changes caused, for instance, by illu- mination, expression or pose changes. Examples of tech- niques from this research direction can be found in [1] or [2]. The second research direction with respect to ro- bust face recognition focuses on combining several facial representations and overcoming externally induced facial appearance changes by considering information in differ- ent feature spaces. Examples of techniques from this di- rection are presented in [3] or [4]. This work has been supported in parts by the national research program P2-0250(C) Metrology and Biometric Systems, the postdoc- toral project BAMBI (ARRS ID Z2-4214), by the European Union, Eu- ropean Regional Fund, in scope of the framework of the Operational Programme for Strengthening Regional Development Potentials for the Period 2007-2013, contract No. 3211-10-000467 (KC Class) and Eu- ropean Union’s Seventh Framework Programme (FP7-SEC-2011.20.6) under grant agreement number 285582 (RESPECT). The authors addi- tionally appreciate the support of COST Actions IC 1106 and IC1206. This paper introduces an novel approach to face recog- nition that can be classified among the techniques from the second research direction, as it exploits representation plurality to achieve robust recognition. The proposed ap- proach crops the facial region from each input image in three different ways. It then maps each of the three crops into one of four color representations and finally extracts several feature types from each of the twelve facial rep- resentations. As will be shown, the described procedure results in a total of thirty facial representations that are combined at the matching score level using a fusion ap- proach based on linear logistic regression (LLR). The presented approach was recently enlisted 1 among the participants of the 2013 face recognition competi- tion that was held in conjunction with the IAPR Inter- national Conference on Biometrics, which is one of the premier conferences on the subject of biometrics. The competition was conducted on the MOBIO database [5], which features facial images taken in uncontrolled con- ditions using various mobile devices. The enlisted ap- proach achieved the best overall performance among all participants and demonstrated highly robust performan- ce [6]. The goal of this paper is two-fold: i) to describe the system enlisted in the 2013 ICB face recognition com- petition, and, more importantly, ii) to present some gen- eral findings made during our development work. The rest of the paper is structured as follows: Sec- tion 2 introduces the novel plurality-based approach, but adjusts the description to the specifics of the competition. In Section 3 experiments conducted during the develop- ment work are presented. These experiments assess dif- ferent aspects of the proposed approach and are important to the wider face recognition community as well. Com- parative results with other participants are also presented in this section. The paper concludes with some final com- ments in Section 4. 2 Proposed methodology 2.1 Overview Before we start with the description of the novel approach, please note that the eye-coordinates of all facial images were provided with the MOBIO database by the compe- 1 As a representative of the University of Ljubljana and Alpineon d.o.o.