No-Reference Video Shakiness Quality Assessment

Video shakiness is a common problem for videos captured by hand-hold devices. How to evaluate the influence of video shakiness on human perception and design an objective quality assessment model is a challenging problem. In this work, we first conduct subjective experiments and construct a data-set with human scores. Then we extract a set of motion features related to video shakiness based on frequency analysis. Feature selection is applied on the extracted features and an objective model is learned based on the data-set. The experimental results show that the proposed model predicts video shakiness consistently with human perception and it can be applied to evaluating the existing video stabilization methods.