On Human Perception-Guided Video Stability Assessment

Abstract

User experience can be compromised if a video contains an in-capture distortion called camera instability, characterized by a typical visual effect caused by uncontrolled camera motion during recording. The process of removing this distortion is known as video stabilization. Despite recent advances in this field, few studies aim to define how video stability quality should be evaluated. In this work, we propose assessment metrics based on pixel profile derivatives and machine learning regressors. We compare them with existing measures on public datasets, correlating the results with human perception scores. Our findings indicate that a widely used measure, called Low-High Frequency Ratio (LHR), correlates poorly with human perception, achieving PLCC=0.388 on LIVE-Qualcomm and PLCC=0.538 on MIND-VQ. In contrast, the best result obtained in this work achieves PLCC=0.884 on MIND-VQ, improving LHR by 34.6 percentage points on that dataset.