WebRTC based video quality assessment

Abstract

The video quality assessment (VQA) methods have been the default mechanism to evaluate the quality of a video. Notably for online video consumption, the video quality might be inconsistent due to many variables that could affect the quality of experience (QOE) for the end user. Therefore, a VQA metric is important to provide a better QOE for the end user. One of the metrics that can be used to determine the video quality is temporal masking. Temporal masking has been researched on since the early days and is a technique to delude individuals into thinking they can identify any differences with a complex spatial and temporal background. Research have been done to evaluate the spatio-temporal in a video using 3D CNN and the results are better compared to other state-of-the-arts technology. In this paper, we aim to explore how Transformers model will perform compared to traditional method of using Convolutional Neural Network (CNN). Transformers have recently started to show effectiveness in solving computer vision tasks. The Transformers' capability to model long-range relationships makes it an obvious choice for learning temporal information over numerous frames for video comprehension problems.