An Omnidirectional Videos Quality Assessment Method Using Salient Object Information

Omnidirectional videos(ODVs) aim to provide an immersive viewing experience but are susceptible to various distortions during processing, leading to quality degradation. Accurately predicting their quality is crucial for ensuring a satisfying users’ Quality of Experience (QoE). While watching ODVs, users’ visible areas are confined to specific portions, referred to as viewports. The choice of a viewport extraction method significantly influences evaluation results, yet existing methods often underutilize advanced semantic information, such as salient objects in videos. In response to this limitation, this study introduces a two-stage model comprising a viewport extraction stage and a quality assessment stage. During the viewport extracion stage, we first employ a viewport extraction network to generate multiple candidate viewports. Subsequently, we introduces an innovative non-maximum suppression algorithm based on salient objects (SO-NMS) to filter the recommended viewports. In the quality assessment stage, we utilizes a dense network to predict the quality scores of each viewport and then integrates these scores to obtain the video quality score. Experimental results demonstrate that the proposed model exhibits exceptional performance, surpassing the majority of existing omnidirectional videos objective quality assessment models.