Unsupervised learning framework for interest point detection and description via properties optimization. (April 2021)

• Record Type: Journal Article
• Authors: Yan, Pei; Tan, Yihua; Tai, Yuan; Wu, Dongrui; Luo, Hanbin; Hao, Xiaolong
• Is Part Of: Pattern recognition. Volume 112(2021)
• Publication Date: 2021-04
• Subjects: Interest point -- Unsupervised learning -- Expectation maximization -- Properties -- Convolution neural network; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2020.107808
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: Unified formulation with latent variable jointly optimizes different properties. Informativeness property encourages features to be extracted in complicated area. Approximate EM algorithm efficiently maximizes the non-differentiable objective. Abstract: This paper presents an unsupervised interest point detection and description method named Properties Optimization Point (POP), which provides a unified objective to optimize different properties of interest point. First, the proposed objective formulates the interest point set as a latent variable, which is flexible to integrate different properties. With the latent variable, the probability formulations are designed for four traditional properties (sparsity, repeatability, invariability, discriminability). Second, a novel property termed as informativeness indicating the information complexity of a local area is designed to determine the areas containing high information, from which interest points are encouraged to be extracted. Third, an efficient approximate Expectation Maximization is proposed to optimize the non-differentiable objective which integrates the above five properties. Finally, POP is instantiated with fully convolutional networks. Experimental results demonstrate that POP outperforms state-of-the-art methods on a number of image matching benchmarks containing both planar and non-planar scenes.