Quality Index for View Synthesis by Measuring Instance Degradation and Global Appearance
Leida Li,Yu Zhou,Jinjian Wu,Fu Li,Guangming Shi
DOI: https://doi.org/10.1109/tmm.2020.2980185
IF: 7.3
2021-01-01
IEEE Transactions on Multimedia
Abstract:Virtual view synthesis plays a vital role in the application of multi-view and free-viewpoint videos. Depth-image-based rendering (DIBR) is the most commonly used approach in view synthesis, and many DIBR algorithms have been proposed. However, how to evaluate the quality of DIBR-synthesized images and benchmark the DIBR algorithms are still very challenging, which may hinder the further development of the view synthesis technique. Hence, an effective quality metric for evaluating the distortions in view synthesis is urgently needed. With this motivation, this paper presents a quality index for view synthesis by simultaneously measuring local Instance DEgradation and global Appearance (IDEA). Due to the imperfection of rendering algorithms, local geometric distortions are easily introduced around instance contours, causing instance degradation, which is the dominant distortion in synthesized views. In this work, image instances are first detected and local instance degradation is measured based on discrete orthogonal moments. Meantime, we propose to measure the global appearance of synthesized images based on the superpixel representation. By integrating both local and global aspects of the distortions, a more accurate quality model is built for view synthesis. Extensive experiments and comparisons have demonstrated the superiority of the proposed method in evaluating the quality of DIBR-synthesized images and benchmarking the performance of view synthesis algorithms.
computer science, information systems,telecommunications, software engineering