MV-MAP: multiresolution video visualization and summarization on MAPs

Abstract

This paper considers visualizing and summarizing image sequences using manifold learning and multiresolution techniques. The images in a video are found usually lying on a significantly low-dimensional manifold, which provides intrinsic information on the video content and formation. The parameterization of the manifold is discovered using a nonlinear subspace method preserving underlying geometry, especially local topology, in the original space. Two modes of video roadmaps have been constructed using VMAPs. The first discovers the landmark points signaling dramatic changes in video content in the temporal order. The second reveals the global content coherence, without the temporal ordering. To facilitate the browsing of long sequences with complicated contents and structures, we build multiresolution visualization and summarization tools on VMAPs. Experimental results validate the proposed method. It may find applications to video monitoring and surveillance for interactive exploitation of video contents, intrusion detection, etc.