Weakly supervised landmark labeling in searched data

  • Authors:

  • Yanyun Qu;Cheng Chen;Yanyun Cheng;Zejian Yuan

  • Affiliations:

  • Xiamen University, Fujian Province, P.R.China;Xiamen University, Fujian Province, P.R.China;Xiamen University, Fujian Province, P.R.China;Xi'an Jiaotong University, Shanxi Province, P.R.China

  • Venue:
  • ICIMCS '10 Proceedings of the Second International Conference on Internet Multimedia Computing and Service
  • Year:
  • 2010

Quantified Score

Hi-index 0.00

Visualization

Abstract

Millions of place-specified photos are uploaded on the internet. Modeling and representing the landmark is very important for landmark retrieval and auto-annotation. In this paper, we aim at collecting images with a specific landmark and labeling the landmark in the images. We propose a weakly supervised labeling approach based on both textual and visual features. Firstly, we cluster the raw landmark dataset with noisy annotation according to the visual context, and obtain a summary of the landmark. Secondly, we mine the local features according to the confidence values of key points and locate the initial region of the landmark. Thirdly, we use Multiple Instance Learning to label the landmark. The contribution of this paper is to automatically learn a landmark representation from the searched data based on a weakly supervised learning approach. We implement our approach on "Statue of Liberty", "Bird Nest" and "Notre Dame de Paris". The experimental results demonstrate that our approach is promising.