This paper presents a novel algorithm for automatically detecting global shakiness in casual videos. Per-frame amplitude is computed by the geometry of motion, based on the kinematic model defined by inter-frame geome...This paper presents a novel algorithm for automatically detecting global shakiness in casual videos. Per-frame amplitude is computed by the geometry of motion, based on the kinematic model defined by inter-frame geometric transformations. Inspired by motion perception, we investigate the just-noticeable amplitude of shaky motion perceived by the human visual system. Then, we use the thresholding contrast strategy on the statistics of per-frame amplitudes to determine the occurrence of perceived shakiness. For testing the detection accuracy, a dataset of video clips is constructed with manual shakiness label as the ground truth. The experiments demonstrate that our algorithm can obtain good detection accuracy that is in concordance with subjective judgement on the videos in the dataset.展开更多
基金This work was partially supported by the National Natural Science Foundation of China under Grant No. 61602015, the Open Funding Project of State Key Laboratory of Virtual Reality Technology and Systems at Beihang University under Grant No. BUAAVR-16KF-06, Beijing Natural Science Foundation under Grant No. 4162019, and the Research Foundation for Young Scholars of Beijing Technology and Business University.
文摘This paper presents a novel algorithm for automatically detecting global shakiness in casual videos. Per-frame amplitude is computed by the geometry of motion, based on the kinematic model defined by inter-frame geometric transformations. Inspired by motion perception, we investigate the just-noticeable amplitude of shaky motion perceived by the human visual system. Then, we use the thresholding contrast strategy on the statistics of per-frame amplitudes to determine the occurrence of perceived shakiness. For testing the detection accuracy, a dataset of video clips is constructed with manual shakiness label as the ground truth. The experiments demonstrate that our algorithm can obtain good detection accuracy that is in concordance with subjective judgement on the videos in the dataset.
