Fast and Accurate Face Recognition with Image Sets

For large-scale face recognition applications using image sets, the images of the query set typically lie in compact regions surrounded by a diffuse sea of images of the gallery set. In this study, we propose a fast and accurate method to approximate the distances from gallery images to the region spanned by the query set for large-scale applications. To this end, we propose a new polyhedral conic classifier that will enable us to compute those distances efficiently by using simple dot products. We also introduce one-class formulation of the proposed classifier that can use query set examples only. This makes the method ideal for real-time applications since testing time approximately becomes the independent of the size of the gallery set. One-class formulation is very important for large-scale face recognition problems in the sense that it can be used in a cascade system with more complex and time-consuming methods to return the most promising candidate gallery sets in the first stage of the cascade so that more complex methods can be run on those a few candidate sets. As a result, we strongly believe that the proposed method will impact future methods and it will enable to introduce face recognition methods working in real-time even for large-scale set based recognition problems. Experimental results on both small and moderate sized face recognition datasets support these claims and demonstrate the efficacy of the proposed method. More precisely, the proposed methods achieve the best accuracies on all tested datasets and we obtained improvements around 18% compared to the best performing rival methods on larger datasets.