Subjective Versus Objective Face Image Quality Evaluation For Face Recognition
Original version
10.1145/3345336.3345338Abstract
The performance of any face recognition system gets affected by the quality of the probe and the reference images. Rejecting or recapturing images with low-quality can improve the overall performance of the biometric system. There are many statistical as well as learning-based methods that provide quality scores given an image for the task of face recognition.
In this study, we take a different approach by asking 26 participants to provide subjective quality scores that represent the ease of recognizing the face on the images from a smartphone based face image dataset. These scores are then compared to measures implemented from ISO/IEC TR 29794-5. We observe that the subjective scores outperform the implemented objective scores while having a low correlation with them. Furthermore, we analyze the effect of pose, illumination, and distance on face recognition similarity scores as well as the generated mean opinion scores.