2d Face Recognition: an Experimental and Reproducible Research Survey

Due to its wide range of applications, automatic face recognition is a research area with high popularity. Many different face recognition algorithms have been proposed in the last decades. Nearly every day there is a new face recognition paper sent to a conference or a journal. Often, researchers provide results that rely on a hand-made non-standard evaluation protocol and that are, hence, incomparable to state-of-theart algorithms. Additionally, the source code for the algorithms is often not provided by the researchers. In consequence, face recognition survey papers can only report the results of other papers. In this paper we provide to our best knowledge the first experimental and evaluative study of a variety of state-of-the-art face recognition algorithms that solely relies on open source software, including color-based linear discriminant analysis, local Gabor binary pattern histogram sequence, Gabor graphs using a Gabor-phase based similarity measure and inter-session variability modeling. Together with this paper we supply the source code to re-run all the experiments that we execute in this study. Experiments are performed on many freely available image databases, always following the evaluation protocols that are attached to them. First, we optimize the parameters of all tested algorithms on a single database. This includes finding the best image preprocessing for each algorithm. Then, we test the algorithms against facial variations as expressions, pose and occlusions using the Multi-PIE and the AR face database. Finally, we report the results of these algorithms on CAS-PEAL, MOBIO, SC face, GBU, FRGC and LFW and discuss some other properties of the algorithms. The results show several trends, partially supporting and partially contradicting prevailing beliefs of the face recognition society. First, Gabor wavelet based algorithms perform better than algorithms relying on raw pixel values, and incorporating Gabor phases improves performance; second, color is an important cue for face recognition; third, the inter-session variability modeling algorithm can handle variations in facial expression and partial occlusions best; fourth, if more than one image is provided at enrollment or probe time, algorithms increase performance; and fifth, biased evaluation protocols as in FRGC or CAS-PEAL favor algorithms that make use of identity information at training time, such as linear discriminant analysis and inter-session variability modeling.