Applications

The aim of this project is developing algorithms to correctly estimate camera motion from a given video sequence, by estimating transformations between two consecutive frames. These predicted transformations will then be used to stabilized a video with high frequency to-and-fro motion. Our developed algorithm is based on deep learning, owing to their high representational power. Recent deep learning methods have shown state-of-the-art performance in both homography estimation and video stabilization tasks, but none of the methods specifically handle our application where the two-and-fromotion is high due to a close range camera and also deliberate unlike sudden shakes for videos shot using hand-held cameras. Our method gives promising results so far, with more improvement expected along the way by rigorous testing of more hypotheses, benchmarking e.t.c. We will also compare our method will existing methods, which requires more resources to complete in a given time-frame.