Open Set Face Recognition


Open set recognition is significant for the applications of face recognition system. Prior to closed set recognition, open set recognition makes no assumption that all probes (test exemplars) have been registered on the gallery set (known identities database). An open set face recognition system has to decide whether the probes are known identities or imposters. Imposters are rejected, while genuine identities are accepted and then to be classified. Open set recognition is more practical for applications that usually confronts with unknown people. However, open set recognition is a definitely challenge for modern pattern recognition. Compared with closed set recognition, open set recognition system has to deal with two more issues of rejecting genuine identities or accepting imposters respectively. It is hard to deal with these two issues at the same time for their contradictory characteristics.

In this paper, we propose a novel method to address open set face recognition problem. It extends the general Adaboost face recognition (GAFR) method used for closed set task, and makes it suitable for open set task. Because
of the trade-off between CCR and FAR, there is a performance bottleneck of open set recognition system. Since threshold plays a role of balancing CCR and FAR, just choosing an optimal threshold is not the best way to make a great improvement to recognition performance. Actually, no matter how to choose the threshold, the performance of recognition algorithm is still influenced by the bottleneck. Taking into account that the errors of genuine exemplars and imposters are mainly caused by the overlap between exemplars similarity distribution, it is an essence way to reduce the overlap area, which will reduce FAR without reducing CCR at the same time and achieve a good recognition result. Inspired by the characteristic of Adaboost method, we combine GAFR method with geometric transformation, and realize an obvious reduction in the similarity overlap.

Two Stage Recognition Structure

Comparison with GAFR



Similarity distribution of positive and negative pairs using: (a) GAFR method, (b) our method.


[1] D.Liu, J. Dai and J. Su. Open Set Face Recognition using Adaboost and Genetric Transformation. TR-SJTU-RCIR, October 2008.
[2] D.Liu and J. Dai. Open Set Face Recognition Based on Adaboost. TR-SJTU-RCIR (in Chinese), January 2009.

Back to Projects Page