Automatic Image Annotation for CBIR
Prof Zhou Jie, Tsinghua University, China

• Research problem/goals
  To get annotated labels of the image automatically, and combine detected results into a high level genre classification result like winter, summer, people group or landscape.
• Description
  Image annotation and genre classification based on low-level region location and mid-level semantic term detection is very important for CBIR. This group proposed a unified algorithm for low level feature extraction and scenery detection. They designed and implemented three different series of semantic term detectors, and finally built a demo to show individual result of each mid level detector, as well as high level genre classification result of the image.
• Publication/public impact
  Two papers published: "Bipartite Graph Reinforcement Model for Web Image Annotation" in ACM Multimedia 2007; "A Search-Based Web Image Annotation Method" accepted by IEEE ICME 2007 (Oral), 2007.

Left: Segmentation Result; Mid: Grouping Result;
Right: Detected Human Body Region (black area)

Research on Large-Scale Image Similarity Search
Prof. YU Nenghai, University of Science and Technology of China

• Research problem/goals
  To study the potential applications of large-scale image similarity search.
• Description
  Unlike existing Web 2.0 services, this effort called “Mobile GeoWeb Services” provides location-dependent Web services. For MyFriend service, the group implemented a heterogeneous intercommunication service in which a user can maintain a buddy list, locate friend(s) on a map, and contact friend(s) with the communication vehicle the user most prefers. For MyGuide services, the group has implemented a prototype in which blog data are extracted for geo-entity identification.
• Publication/public impact
  An optimized itinerary-based KNN spatial query in the mobile/wireless sensor environment was developed, and a paper based on the topic was accepted by ACM CIKM 2007. As for MyMusic, the group proposed a clustering-based positioning technique for a large-scale mobile environment, which formed a paper accepted by IEEE MASS 2007.

See more at: http://dmn.ustc.edu.cn/msproject/similarity

Mobile GeoWeb Services: MyFriend, MyGuide and MyMusic
Prof. Peng Wen-Chih; National Chiaotung University, Taiwan

• Research problem/goals
  To develop a way for users to access three location-based Web 2.0 services through use of mobile divisions, namely MyFriend, MyGuide and MyMusic.
• Description
  “Similar” for large-scale image similarity search has two definitions: visually similar and semantically similar. For the first definition, the group proposed to use the compact representation of images to search visually similar images efficiently; for the second one, they proposed to utilize automatic image annotation technology to tag semantically similar images. They also built a more than 11-million images database from the Internet to support current and future experiments and systems, such as CBIR, text-based image search and automatic image annotation.
• Publication/public impact
  Two papers published: “Adaptive skin detection based on online training” and “Rating of web pages based on image content analysis” in Proceedings of SPIE - Multispectral Image Processing and Pattern Recognition, 2007.

System demonstration