Hybrid Unsupervised Scale-invariant Slide Detection (HUSSD) for Video Presentation
Bingzhao Shan, Muhammad Rizwan Abid, Ehsan Amjadian · 2020
This paper addresses the task of unsupervised scale-invariant slide detection for video presentations, which aims to detect slides and index them with an appropriate location in the presentation video. We propose two new methods that improve on Scale-Invariant Feature Transform (SIFT) in terms of computational cost as well as accuracy. Both methods are instantiations of our Hybrid Unsupervised Scale-invariant Slide Detection paradigm (HUSSD). The first is a feature-based HUSSD algorithm. This method utilizes SIFT as well as Oriented FAST and rotated BRIEF (ORB) tracker to resolve the computation cost issue. Feature-based HUSSD optimizes for speed and substantially improves the runtime to approximately 30 times faster than SIFT at the trivial cost of accuracy falling from 90.06% to 82.89%. Furthermore, our second novel method for slide detection, namely template-based HUSSD, employs a multiscale matching detector and a single scale template matching tracker to improve on SIFT in terms of both speed and accuracy by 7.91 times and to 95.05% respectively.