Saliency in objective video quality assessment: What is the ground truth?

摘要

Finding ways to be able to objectively and reliably assess video quality as would be perceived by humans has become a pressing concern in the multimedia community. To enhance the performance of video quality metrics (VQMs), a research trend is to incorporate visual saliency aspects. Existing approaches have focused on utilizing a computational saliency model to improve a VQM. Since saliency models still remain limited in predicting where people look in videos, the benefits of inclusion of saliency in VQMs may heavily depend on the accuracy of the saliency model used. To gain an insight into the actual added value of saliency in VQMs, ground truth saliency obtained from eye-tracking instead of computational saliency is an essential prerequisite. However, collecting eye-tracking data within the context of video quality is confronted with a bias due to the involvement of massive stimulus repetition. In this paper, we introduce a new experimental methodology to alleviate such potential bias and consequently, to be able to deliver reliable intended data. We recorded eye movements from 160 human observers while they freely viewed 160 video stimuli distorted with different distortion types at various degradation levels. We analyse the extent to which ground truth saliency as well as computational saliency actually benefit existing state of the art VQMs. Our dataset opens new challenges for saliency modelling in video quality research and helps better gauge progress in developing saliency-based VQMs.