Dual-stage attention based symmetric framework for stereo video quality assessment

Abstract

The compelling creative capabilities of stereo video have captured the attention of scholars towards its quality. Given the substantial challenge posed by asymmetric distortion in stereoscopic visual perception within the realm of stereoscopic video quality evaluation (SVQA), this study introduces the novel $D^{3}Net$ (Dual Branch, dual-stage Attention, Dual Task) framework for stereoscopic video quality assessment. Leveraging its innovative dual-task architecture, $D^{3}Net$ employs a dual-branch independent prediction mechanism for the left and right views. This approach not only effectively addresses the prevalent issue of asymmetric distortion in stereoscopic videos but also pinpoints which view drags the overall score down. To surmount the limitations of existing models in capturing global detail attention, $D^{3}Net$ incorporates a two-stage distorted attention fusion module. This module enables multi-level fusion of video features at both block and pixel levels, bolstering the model’s attention towards global details and its processing capabilities, consequently enhancing the overall performance of the model. $D^{3}Net$ has exhibited exceptional performance across mainstream and cross-domain datasets, establishing itself as the current state-of-the-art (SOTA) technology.