Predicting performance of face recognition systems: An image characterization approach

Predicting performance of face recognition systems on previously unseen data is very useful for deploying these systems in different places. Different extrinsic and intrinsic factors like illumination, pose, expression, etc. affect matching performance of even the best of face recognition algorithms. This makes it difficult for one to accurately predict how a system will perform at a new deployment location with novel imaging conditions. With this motivation, we present a novel framework to predict performance of face matching systems on unseen data without the need of subject-wise labeling of images typically necessary for evaluations. The framework relies on learning a mapping from a space characterizing imaging conditions to the score space using Multi-dimensional Scaling. Extensive evaluation on the Multi-PIE data using different algorithms demonstrates the usefulness of the prediction framework. Experiments using training data which is completely different from the test data further justifies the use of the proposed approach for the task of performance prediction.