Computer Vision Performance and Image Quality Metrics : A Reciprocal Relation

Abstract

Computer vision algorithms are essential components of many systems in operation today. Predicting the robustness of such algorithms for different visual distortions is a task which can be approached with known image quality measures. We evaluate the impact of several image distortions on object segmentation, tracking and detection, and analyze the predictability of this impact given by image statistics, error parameters and image quality metrics. We observe that existing image quality metrics have shortcomings when predicting the visual quality of virtual or augmented reality scenarios. These shortcomings can be overcome by integrating computer vision approaches into image quality metrics. We thus show that image quality metrics can be used to predict the success of computer vision approaches, and computer vision can be employed to enhance the prediction capability of image quality metrics – a reciprocal relation.