Unsupervised shape discovery using synchronized spectral networks. (September 2017)

• Record Type: Journal Article
• Authors: Cai, Yunliang; Lum, Andrea; Mercado, Ashley; Landis, Mark; Warrington, James; Li, Shuo
• Is Part Of: Pattern recognition. Volume 69(2017:Sep.)
• Page Start: 39
• Page End: 51
• Publication Date: 2017-09
• Subjects: Spectral synchronization -- Joint image matching -- Groupwise segmentation -- Shape discovery; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2017.03.032
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: A new unsupervised shape discovery method using a novel joint foreground/background segmentation and a dense part-part correspondence between all image pairs. A new multiscale spectral synchronization method which jointly align the spectral representations of all input images. A novel unsupervised superpixel-based groupwise co-registration which converts the implicit eigenvector-eigenvector synchronization to superpixel-superpixel dense correspondences. Abstract: Unsupervised discovery and extraction of common shapes from unlabeled images is a fundamental problem in object recognition and has broad applications in practice. However, shape discovery suffers from the lack of consistent matching methods for finding the correspondences between objects with different colors/textures among the input images. In this paper, we propose a novel unsupervised shape discovery method using Synchronized Spectral Network (SSN) which provides automatic part-part correspondences across images. The SSN is spectral graph-based model that encodes the pixel self-similarities of different images in spectral bases, and synchronizes the bases between images to achieve the part-part correspondences. Unlike explicit feature matching, correspondences obtained by spectral synchronization are independent of colors/textures and image modalities. An image network can then be built by spectral correspondences where the common shapes among them can be easily identified and segmented. Our results in multiple shape discovery datasets demonstrate that we outperform the state-of-the-art object/shape discovery methods, providing better segmentations for common shapes.