Saliency Inspired Quality Assessment of Stereoscopic 3D Video

TL;DR

This paper explores how incorporating 3D visual saliency models enhances the accuracy of both full-reference and no-reference quality assessment metrics for stereoscopic 3D videos, based on eye tracking data.

Contribution

It introduces the integration of state-of-the-art 3D visual attention models into existing 3D video quality metrics, demonstrating performance improvements.

Findings

Saliency maps generally improve quality assessment accuracy.

Performance gains vary with distortion type and metric used.

3D VAM integration enhances both FR and NR quality metrics.

Abstract

To study the visual attentional behavior of Human Visual System (HVS) on 3D content, eye tracking experiments are performed and Visual Attention Models (VAMs) are designed. One of the main applications of these VAMs is in quality assessment of 3D video. The usage of 2D VAMs in designing 2D quality metrics is already well explored. This paper investigates the added value of incorporating 3D VAMs into Full-Reference (FR) and No-Reference (NR) quality assessment metrics for stereoscopic 3D video. To this end, state-of-the-art 3D VAMs are integrated to quality assessment pipeline of various existing FR and NR stereoscopic video quality metrics. Performance evaluations using a large scale database of stereoscopic videos with various types of distortions demonstrated that using saliency maps generally improves the performance of the quality assessment task for stereoscopic video. However, depending on the type of distortion, utilized metric, and VAM, the amount of improvement will change.