Digital Video Stabilization: Methods, Datasets, and Evaluation

Abstract

Our thesis addressed digital video stabilization, a process that removes unwanted shakes from videos via software. We performed a thorough review, which resulted in two survey papers. We also studied and proposed a new stability measure aligned with human perception and a novel method for evaluating 2D camera motion to assess video quality better. Next, we introduced NAFT, a semi-online DWS with a new neighborhood-aware mechanism. This method stabilizes videos without relying on an explicit definition of stability. To train NAFT effectively, we created SynthStab, a paired synthetic dataset. NAFT achieves stabilization quality comparable to non-DWS methods, with a significantly smaller model (a 14× reduction).