Two-Level-Granularity Manifold Learning Algorithm for Video Visualization

Visualization of high-dimensional video data is an important problem in computer vision and machine learning. It not only needs to find the low-dimensional intrinsic laws of video set, but also explores the low-dimensional geometric distribution of all frames from videos. This paper presents a new technique called "Two-level-granularity Manifold Learning (TML)" that visualizes high-dimensional video data from two granular levels. Experiments on Cambridge gesture database show that our TML method can obtain better understandable visualization results and has lower time complexity than ISOMAP algorithm.