Soutenance de thèse : Yana NEHME

Visual Quality of Rendered 3D Meshes with Color Attributes: Subjective and Objective Evaluation

Doctorante : Yana NEHME

Laboratoire INSA : LIRIS

Ecole doctorale : ED512 Informatique Et Mathématiques de Lyon

As technological advances and capabilities in the field of computer graphics grow day by day, the need to master the visualization and processing of 3D data increases at an equal pace. Indeed, the development of modeling software and acquisition devices makes 3D graphics rich and realistic: complex models enriched with various appearance attributes. The way this 3D content is consumed is also evolving from standard screens to Virtual and Mixed Reality (VR/MR). However, the size and complexity of these rich 3D models often make their interactive visualization problematic. This is particularly the case in immersive environments and online applications. Thus, to avoid latency and rendering issues, diverse processing operations, including simplification and compression, are usually applied, resulting in a loss of quality in the final rendering. Thus, both subjective studies and objective metrics are needed to predict this visual loss and assess the quality as perceived by human observers. In this thesis, we address the aforementioned challenges. We conduct an extensive study to determine the best subjective quality assessment methodology to adopt for assessing the visual quality of 3D graphics, especially in VR. We establish the two largest (to date) public quality assessment datasets for animated meshes with color attributes in the form of vertex color and texture maps, produced in VR and crowdsourcing respectively. Moreover, we provide an in-depth analysis of the influence of source model characteristics, distortion interactions, viewpoints and animations on the perceived quality of 3D graphics. In this line of research, two novel objective quality assessment metrics are devised for 3D meshes with color attributes. To the best of our knowledge, our proposed metrics are the first attempts to integrate both geometry and color information for quality assessment of such data. Lastly, we investigate how incorporating the saliency into these metrics improves the predicted quality.