Lecturer: Li Hui
Lecture Time: June 7th 4:00 pm
Lecture Place: Room 417 Information Building
DPR-Tree An Efficient Indexing Method for Multimedia kNN Search
kNN search is an important task in many modern multimedia applications. Although there is a long stream of research on high-dimensional multimedia indexing,aiming to efficiently organize multimedia data to facilitate kNN search, most of them are focused on how to reduce the IO operations to improve the overall kNN search performance. In fact, in many scenarios, IO operations are only a minor factor of kNN search overhead. The efficiency of the existing indexing approaches does not always suffice.
In this talk, I will address the multimedia indexing problem from the perspective of computation cost reduction. We devised a multimedia index structure named DPR-Tree, and applied a novel kNN search
approach to reduce the overhead of distance computation and speed up the kNN query processing. We conducted an extensive experimental study on a large real-life multimedia dataset called MSRA-MM 2.0, obtained from the commercial Bing Image search engine. Our results show that the DPR-Tree based high-dimensional multimedia kNN search is significantly faster than other indexing methods without incurring accuracy loss.
