Image Quality Assessment in End-to-end Face Analytics Systems

End-to-end face analytics systems perform tasks such as face detection, alignment, and recognition sequentially, and the performance of each task depends on the success of the preceding tasks. Typically, these systems are deployed in resource and bandwidth-constrained environments. In such systems, resource utilization can be improved by linking the quality of face images processed by these systems to the performance of analytics tasks. This approach ensures that the system processes face images that are best suited for analytics. Recently, the development of face image Quality Estimators (QEs) has received significant attention. However, none of these face image QEs have been evaluated in an end-to-end manner to determine how they affect the overall performance of a face analytics system. In this paper, we utilize a carefully curated dataset to evaluate task-specific face image QEs in the context of an end-to-end face analytics system. Through our experiments, we demonstrate that task-specific QEs are effective in rejecting images not suitable for end-to-end face analytics systems and can significantly improve overall system performance.