Factored Representation of Local Feature Set for Affine Invariant Matching

2018 Fiscal Year Final Research Report

• Project/Area Number: 16K00253
• Research Category: Grant-in-Aid for Scientific Research (C)
• Allocation Type: Multi-year Fund
• Section: 一般
• Research Field: Perceptual information processing
• Research Institution: Chubu University
• Principal Investigator: FUJIYOSHI Hironobu 中部大学, 工学部, 教授 (20333172)
• Co-Investigator(Kenkyū-buntansha): 上瀧 剛 熊本大学, 大学院先端科学研究部(工), 准教授 (20582935)
• Research Collaborator: AMBAI Mitsuru
• Project Period (FY): 2016-04-01 – 2019-03-31
• Keywords: キーポイントマッチング / 特徴記述 / 多視点記述子

Outline of Final Research Achievements

Conventional multi-viewpoint descriptors, such as Affine SIFT (ASIFT), require much online affine-warping of a patch image to precisely match images that have viewpoint changes. Therefore, we propose affine invariant descriptor without conventional heavy online affine-warping. To this end, the proposed descriptor represents traditional local descriptors as the inner product between “feature-description filters” and a local patch image. By using feature-description filters, we can conduct affine-warping efficiently using pre-computed filter sets. In addition, affine-warped filters can be compactly represented using a factorization method, and the multi-viewpoint local features can be generated for arbitrary continuous affine parameters. Experimental results indicate that the proposed descriptor describes multi-viewpoint features more efficiently than conventional affine invariant descriptors while maintaining the keypoint matching performance.

Free Research Field

画像認識

Academic Significance and Societal Importance of the Research Achievements

撮影位置を不連続にして撮影した画像には，視点変化によって生じる変化が含まれる．このような画像をつなぎ合わせるには，視点変化に影響を受けない特徴量記述が必要となる．本研究成果はこの問題を解決するアプローチであり，従来法より大幅な高速化を実現したことで，携帯型デバイスにおいての実行が可能となる．これにより，画像認識技術を用いた様々なアプリケーションの開発に貢献できる．