Full Reference Quality Assessment of DIBR-Based Synthesized Images

Abstract

Nowadays, free viewpoint video applications provide more immersive, interactive and appealing experiences to the end user. Usually, a Multiview plus Depth (MVD) 3D representation model is used, where only a few views are transmitted and the rest are synthesized at the receiver, often with a Depth Image Based Rendering (DIBR) technique. However, artifacts resulting from errors introduced by the view synthesis solution (often in occluded areas), as well as from the compression of the texture and depth, may degrade the synthesized view quality; to automatically evaluate the resulting quality, objective quality assessment metrics for synthesized views are then required. This paper proposes a full reference quality assessment metric for synthesized images; it seeks to detect and quantify edges distortions, where synthesis errors significantly impact the perceived quality, using a new variant of the Hausdorff distance. The metric was evaluated with a public database, and compared with conventional 2D and synthesized images quality assessment metrics. The proposed metric provides a good prediction of the impact that synthesis artifacts have on the perceived image quality, outperforming the considered state-of-the-art metrics.