Blind Video Stabilization Quality Assessment based on convolutional LSTM

Assessing the performance of video stabilization algorithms is still in its infancy compared with the state-of-the-art in video quality assessment, where the focus is on classic distortions such as noise, blur, lighting problems, and coding artifacts. In this study, we present a new blind video stabilization quality assessment metric based on learning sensor outputs via a novel deep neural network. We build an architecture based on a convolutional LSTM to estimate accelerations given by an inertial measurement unit sensor with no additional information than video frames. The experimental study was carried out on a new dedicated dataset, which showed the effectiveness of our metric in characterizing the level of shakiness of videos, in addition, we demonstrate the efficiency of our system to learn acceleration data provided from sensors, which opens ambitious perspectives in retrieving sensors data from any video.