Designing from Motivation: Exploring Large Scale Tagged Data Collection through Social Monetization Computing

  • Author / Creator
    Wu, Qiong
  • With the exponential growth of web image data, image tagging is becoming crucial in many image based applications such as object recognition and content-based image retrieval. However, despite the great progress achieved in automatic recognition technologies, none has yet provided a satisfactory solution to be widely useful in solving generic image recognition problems. Automatic technologies usually make certain assumptions, such as a limited number of object categories and how many objects there are in an image. With the goal of tagging generic images, so far, only manual tagging can provide precise image descriptions. However, the cost and tediousness of manual tagging is the major concern. The first effort to motivate people to tag images is the ESP game, proposed by Luis von Ahun. In the same vein, we ask the same question how can we motivate people to tag web images, which belongs to the research field of collective intelligence. So far, crowdsourcing, human computation (ESP game) and social computing are three major methods resolving the problem of motivating people to work collaboratively and to produce something intelligent. However, none of them can achieve the goal of collecting large scale tagged images at high quality for low cost. In this thesis, we propose a Social Monetization Computing (SMC) model, which incorporates monetary incentives into social computing to guarantee high quality work from both crowdsourcing workers and social web users for a low cost. In addition, we summarize a design guidance of a SMC system. In the light of SMC system design guidelines, we describe the evolutionary design and implementation of an image tagging system, called EyeDentifyIt, driven by image-click-ads framework. A series of usability studies are presented to demonstrate how EyeDentifyIt provides better user motivations, produces higher quality data, and requires less workload from workers, compared to state-of-the-art approaches. To further reduce workload involved in the image tagging process, we develop an efficient method for automatically parsing fashion images, which resolves three common problems including occlusions, background spills and over smoothing of infrequent labels, in existing fashion parsing methods. The experiment results demonstrate that the proposed method outperforms state-of-the-art clothing parsing methods from both quantity and quality perspectives.

  • Subjects / Keywords
  • Graduation date
  • Type of Item
  • Degree
    Doctor of Philosophy
  • DOI
  • License
    This thesis is made available by the University of Alberta Libraries with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.
  • Language
  • Institution
    University of Alberta
  • Degree level
  • Department
    • Department of Computing Science
  • Supervisor / co-supervisor and their department(s)
    • Boulanger, Pierre (Computing Science)
  • Examining committee members and their departments
    • Cheng, Irene (Computing Science)
    • Zhao, Vicky H (Electrical and Computer Engineering)
    • Hindle, Abram (Computing Science)
    • Huang, Xiaolei (Computer Science & Engineering)
    • Boulanger, Pierre (Computing Science)