Implementation and Analysis of a Low-Latency Off-Loading Approach to Object Detection for Augmented Reality on a Mobile Phone
Term of Award
Master of Science in Computer Science (M.S.)
Document Type and Release Option
Thesis (restricted to Georgia Southern)
Copyright Statement / License for Reuse
This work is licensed under a Creative Commons Attribution 4.0 License.
Department of Computer Science
Committee Member 1
Committee Member 2
This paper presents a mobile augmented reality system (AR) that utilizes advancements in object recognition, specifically a convolutional neural network (CNN), to accurately annotate points of interest while maintaining imperceptible latency. Provided in this thesis is a detailed analysis of how this mobile AR system achieves this low latency. This detailed analysis includes how the offloading pipeline was optimized to reduce end-to-end latency in the system without sacrificing the CNN's accuracy. Additionally, another optimization was introduced in the form of a simple metric that is employed to determine the quality of an image captured for the purposes of reducing wasted computation by offloading images of inferior quality. How this system scales with an increasing number of users is also explored with the help of Nvidia's Multi-Process Service (MPS). To test this system, a campus tour application is made which annotates places of interest on Georgia Southern's campus.
Maner, Scott, "Implementation and Analysis of a Low-Latency Off-Loading Approach to Object Detection for Augmented Reality on a Mobile Phone" (2019). Electronic Theses and Dissertations. 1986.
Research Data and Supplementary Material