Vega: A Computer Vision Processing Enhancement Framework with Graph-based Acceleration
Files
Date
2020-01-07
Authors
Contributor
Advisor
Department
Instructor
Depositor
Speaker
Researcher
Consultant
Interviewer
Narrator
Transcriber
Annotator
Journal Title
Journal ISSN
Volume Title
Publisher
Volume
Number/Issue
Starting Page
Ending Page
Alternative Title
Abstract
The popularity of Computer Vision (CV) algorithms has been on the rise given their growing dependence on machine learning and deep neural networks. The resulting improvement in inference accuracy has revolutionized a number of fields. However, given that CV algorithms consist of many different stages, each having different computing characteristics, their execution is frequently irregular and inefficient, unable to leverage the full potential of the computing platform. Presently, supporting real-time video processing for high resolution images on edge systems involves a significant amount of programming effort and performance tuning. To overcome this challenge, we present Vega, a parallel graph-based framework that enables better utilization of multi-core edge computing platforms. Vega provides a highly flexible and user-friendly interface to execute a range of CV algorithms efficiently, leveraging multiple external libraries for performance. First, Vega maps independent stages of a CV algorithm to nodes in a pipeline graph. Next, it dynamically schedules nodes on a multi-core CPU using multi-threading. From our experimental results, our framework improves performance of all selected algorithms by at least 1.75x and up to 4.82x on the same platform. We analyze the impact of using our framework in terms of hardware utilization, frame processing latency and throughput.
Description
Keywords
Systems and Applications of Video Analytics, acceleration, computer vision, face detection, image processing
Citation
Extent
10 pages
Format
Geographic Location
Time Period
Related To
Proceedings of the 53rd Hawaii International Conference on System Sciences
Related To (URI)
Table of Contents
Rights
Attribution-NonCommercial-NoDerivatives 4.0 International
Rights Holder
Local Contexts
Collections
Email libraryada-l@lists.hawaii.edu if you need this content in ADA-compliant format.