Computational Analysis of Degradation Modeling in Blind Panoramic Image Quality Assessment

TL;DR

This work tackles the easy-database issue in blind panoramic image quality assessment (BPIQA) by analyzing how dataset content and distortion diversity affect benchmarking and by proposing a generic framework that mimics human viewing with four modular components: a viewport generator, a viewport-wise feature extractor, an inter-viewport interaction module, and a quality regression module. Through extensive cross-database experiments on eight PIQA datasets, the study demonstrates that easy databases can shrink the gap between BIQA and BPIQA, cause saturation, and obscure the benefits of specialized designs, whereas diverse heterogenous-distortion data (e.g., OIQ-10K, JUFE-10K) improve generalization. The findings show transformer-based and hybrid architectures offer advantages for panoramic-specific distortions and emphasize the need for more representative, hard-sample datasets to drive robust BPIQA. Overall, the paper provides methodological guidance and empirical evidence to advance BPIQA toward reliable, real-world performance.

Abstract

Blind panoramic image quality assessment (BPIQA) has recently brought new challenge to the visual quality community, due to the complex interaction between immersive content and human behavior. Although many efforts have been made to advance BPIQA from both conducting psychophysical experiments and designing performance-driven objective algorithms, \textit{limited content} and \textit{few samples} in those closed sets inevitably would result in shaky conclusions, thereby hindering the development of BPIQA, we refer to it as the \textit{easy-database} issue. In this paper, we present a sufficient computational analysis of degradation modeling in BPIQA to thoroughly explore the \textit{easy-database issue}, where we carefully design three types of experiments via investigating the gap between BPIQA and blind image quality assessment (BIQA), the necessity of specific design in BPIQA models, and the generalization ability of BPIQA models. From extensive experiments, we find that easy databases narrow the gap between the performance of BPIQA and BIQA models, which is unconducive to the development of BPIQA. And the easy databases make the BPIQA models be closed to saturation, therefore the effectiveness of the associated specific designs can not be well verified. Besides, the BPIQA models trained on our recently proposed databases with complicated degradation show better generalization ability. Thus, we believe that much more efforts are highly desired to put into BPIQA from both subjective viewpoint and objective viewpoint.

Figures (3)

• Figure 1: An illustration of the advance of PIQA in terms of pearson linear correlation coefficient on the CVIQ database sun2018large. The listed PIQA models include WS-SSIM zhou2018weighted, MC360IQA sun2019mc360iqa, VGCN xu2020blind, Zhou21 zhou2021omnidirectional, Liu21 liu2021blind, SAL-360IQA sendjasni2022sal, AHGCN fu2022adaptive, SSS liu2022hvs, Zhou23 zhou2023perception, ST360IQ tofighi2023st360iq, and Assessor360 wu2023assessor360. Note that the phenomenon of performance saturation on the CVIQ database also suits other PIQA databases.
• Figure 2: The visual examples of (a) homogeneously distorted panoramic image from OIQA database duan2018perceptual, (b) heterogeneously distorted panoramic image from OIQ-10K database yan2024omnidirectional and (c) generated panoramic image from AIGCOIQA database yang2024aigcoiqa2024. Note that viewports outlined in green indicate high visual quality, while those with red contours signify low visual quality.
• Figure 3: The analysis framework for investigating the easy-database issue. We carry out this study from three perspectives: 1) the gap between BPIQA and BIQA (subfig (a)), 2) the necessity of specific design in BPIQA models (subfig (b)), and 3) the generalization ability of BPIQA models (subfigs (c) and (d)).