An ‘Unsupervised’ 3D Image Quality Assessment using Spectral Decompositions of Scene Components

We present an unsupervised no-reference (NR) image quality assessment (IQA) algorithm designed to evaluate perceptual quality in natural stereoscopic 3D (S3D) scenes. At the core of our approach is the creation of a cyclopean image - a synthesized single view that integrates information from both left and right perspectives to simulate human binocular perception. This cyclopean image serves as the foundation of our quality assessment method, closely capturing the chrominance, structural and depth characteristics that influence perceived S3D quality. To further enhance the analysis, we apply a multi-orientation steerable pyramid decomposition on the cyclopean image, capturing intricate chrominance, luminance and depth-related features through entropy and Perception-based Image Quality Evaluator scores of each orientation. These orientation-specific scores are then aggregated to form a comprehensive quality metric for the S3D image. The proposed model, an entirely unsupervised model that operates without reference images or prior training on subjective opinion scores, demonstrates consistent performance across various distortion types in the LIVE Phase I and Phase II stereoscopic datasets.