Blind S-3D VR picture quality prediction using trivariate brightness, color, and disparity statistics

Abstract

We introduce a new supervised no-reference stereoscopic (3D or S3D) virtual reality (VR) image quality assessment (IQA) model that is based on measuring joint statistical dependencies between luminance, chroma, and disparity cues. To accomplish this a Trivariate generalized Gaussian distribution (TGGD) model is deployed to capture the relationships between these three dependent sources of visual information. The TGGD model is evaluated at multiple scales and orientations of steerable pyramid decompositions of cube map projected (CMP) faces computed on S3D-VR views. The model parameters (shape and spread) of the TGGD, along with measures of dissimilarity and fractional anisotropy are estimated using the eigenvalues of the TGGD covariance matrix. The computed features are consolidated using a weighting mechanism whereby the weights of the CMP face from the left and right S3D-VR views are obtained based on the subband strengths of a saliency map. Further, the energy of the gray level co-occurrence matrix is computed to capture spatial transitions between the CMP faces of the S3D-VR views. The consolidated feature vector set and the human assessment scores are applied as the inputs to an AdaBoost-Back propagation neural network, which learns to predict the perceptual quality of S3D-VR images. We utilize the benchmark LIVE S3D-VR IQA and SOLID S3D-VR IQA datasets to evaluate the efficacy of the proposed model, which we call Tri3D VR-QA, against the state-of-the-art. Experiments on the benchmark LIVE S3D-VR and SOLID S3D-VR datasets demonstrate that Tri3D VR-QA achieves superior correlations with human perception, attaining linear correlation coefficient/Spearman rank order correlation coefficient values of 0.8215/0.8140 and 0.9248/0.9063, respectively, outperforming leading 2D and 3D IQA models by up to 3%–5%. The source code of the proposed Tri3D VR-QA model is publicly available on Google Drive. © 2026 Elsevier B.V. All rights are reserved, including those for text and data mining, AI training, and similar technologies.