Large Scale Nearest Neighbor Search by Acceleration an Stabilization of Product Quantization

2016 Fiscal Year Final Research Report

• Project/Area Number: 15K12025
• Research Category: Grant-in-Aid for Challenging Exploratory Research
• Allocation Type: Multi-year Fund
• Research Field: Multimedia database
• Research Institution: The University of Tokyo
• Principal Investigator: Aizawa Kiyoharu 東京大学, 大学院情報理工学系研究科, 教授 (20192453)
• Project Period (FY): 2015-04-01 – 2017-03-31
• Keywords: 直積量子化 / 最近傍探索

Outline of Final Research Achievements

Product Quantization (PQ) is the most promising technique for high-speed nearest neighbor search for large-scale images retrieval. For this product quantization, we carried out the following research in order to dramatically increase computational efficiency and memory efficiency.
(1)Dense space partitioning PQ: Efficient dense space partitioning by combining multiple centroids of clusters　(2)PQTable: Efficient product quantization using PQ code as a hash table　(3)PQkmeans: Fast and memory efficient kmeans in PQ domain achieving very large scale clustering　(4)Residual Expansion Algorithm: Efficient and Effective Optimization for non-convex least square problem such as kmeans

Free Research Field

マルチメディア、画像処理、コンピュータビジョン