Dual Dimensionality Reduction for Efficient Video Similarity Search

Abstract

With ever more advanced development in video devices and their extensive usages, searching videos of user interests from large scale repositories, such as web video databases, is gaining its importance. However, the huge complexity of video data, caused by high dimensionality of frames (or feature dimensionality) and large number of frames (or sequence dimensionality), prevents existing content-based search engines from using large video databases. Hence, dimensionality reduction on the data turns out to be most promising. In this paper, we propose a novel video reduction method called Optimal Dual Dimensionality Reduction (ODDR) to dramatically reduce the video data complexity for accurate and quick search, by reducing the dimensionality of both feature vector and sequence. For a video sequence, ODDR first maps each high dimensional frame into a single dimensional value, followed by further reducing the sequence into a low dimensional space. As a result, ODDR approximates each long and high dimensional video sequence into a low dimensional vector. A new similarity function is also proposed to effectively measure the relevance between two video sequences in the reduced space. Our experiments demonstrate the effectiveness of ODDR and its gain on efficiency by several orders of magnitude.