Multi-Core Software

Accelerating Video Feature Extractions in CBVIR on Multi-Core Systems

INTRODUCTION

Nowadays, with advances in video capture and storage techniques, the sheer amount of video data has exploded not only in enterprises but also in our homes. Concomitantly, there is an increasing demand for a system that can help end users to index massive amounts of video data for further search, browse, and management tasks. Digital home-usage media centers are coming into being for this very purpose. Most of these centers consist of two key ingredients: the Content-Based Video Information Retrieval (CBVIR) module and the computing platform.

CBVIR is a computational technique to index unstructured video information in terms of low-level audio/visual features [1]. MPEG-7 is an experimental standard acting as a guideline for low-level audio/visual feature extractions [2]. It includes a set of visual color, texture, shape, and motion descriptors. Since low-level visual feature extraction is the most time-consuming part in CBVIR applications, these applications are much more compute intensive than traditional video decoding/encoding applications. Although typically the indexing can be done in off-line mode, there are many more emerging scenarios that require a real-time or even super-real-time processing capability in a CBVIR system. With the boom in multi-core processors, we can take full advantage of the computing power of today's multi-core platform to accelerate the use of CBVIR applications [3].

In this paper, we optimize and parallelize a set of typical feature extraction applications on a multi-core system. Our results show most of them are much slower than real-time in their original implementations. After serial optimization, however, they become 3.3x faster, and only five of them are still slower than real-time. After the tailored parallelization, the six most compute-intensive applications obtain up to a 7.6x speedup on a dual-socket, quad-core system, which enables them to achieve super-real-time performance.

This paper is organized as follows. First, we briefly review several low-level visual descriptors under the guidelines of the MPEG-7 experimental standard. Next, we present our optimization and parallelization methodology for low-level visual feature extractions. Then, we show the performance analysis results of the typical feature extraction workloads.

Back to Top