Deep unsupervised similarity learning using partially ordered sets

• Verfasst von: Bautista, Miguel [VerfasserIn]
• Verfasst von: Sanakoyeu, Artsiom [VerfasserIn]
• Verfasst von: Ommer, Björn [VerfasserIn]
• Titel: Deep unsupervised similarity learning using partially ordered sets
• Verf.angabe: Miguel A. Bautista, Artsiom Sanakoyeu, Björn Ommer
• E-Jahr: 2017
• Jahr: 09 November 2017
• Umfang: 10 S.
• Teil: year:2017
• Teil: pages:1923-1932
• Teil: extent:10
• Fussnoten: Gesehen am 13.02.2018
• Titel Quelle: Enthalten in: IEEE Conference on Computer Vision and Pattern Recognition (30. : 2016 : Honolulu, Hawaii)2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR)
• Ort Quelle: Piscataway, NJ : IEEE, 2017
• Jahr Quelle: 2017
• Band/Heft Quelle: (2017), Seite 1923-1932
• DOI: doi:10.1109/CVPR.2017.208
• Datenträger: Online-Ressource
• Sprache: eng
• Sach-SW: CNN
• Sach-SW: compact surrogate classes
• Sach-SW: Computational modeling
• Sach-SW: Computer vision
• Sach-SW: deep unsupervised similarity learning
• Sach-SW: detailed pose estimation
• Sach-SW: fine-grained similarities
• Sach-SW: grouping procedure
• Sach-SW: image classification
• Sach-SW: learning (artificial intelligence)
• Sach-SW: local partial orders
• Sach-SW: neural nets
• Sach-SW: object classification
• Sach-SW: partial ordering task
• Sach-SW: partially ordered sets
• Sach-SW: pose estimation
• Sach-SW: similarity learning
• Sach-SW: training data
• Sach-SW: unsupervised approach
• Sach-SW: Visualization
• K10plus-PPN: 156978292X

Abstract

Unsupervised learning of visual similarities is of paramount importance to computer vision, particularly due to lacking training data for fine-grained similarities. Deep learning of similarities is often based on relationships between pairs or triplets of samples. Many of these relations are unreliable and mutually contradicting, implying inconsistencies when trained without supervision information that relates different tuples or triplets to each other. To overcome this problem, we use local estimates of reliable (dis-)similarities to initially group samples into compact surrogate classes and use local partial orders of samples to classes to link classes to each other. Similarity learning is then formulated as a partial ordering task with soft correspondences of all samples to classes. Adopting a strategy of self-supervision, a CNN is trained to optimally represent samples in a mutually consistent manner while updating the classes. The similarity learning and grouping procedure are integrated in a single model and optimized jointly. The proposed unsupervised approach shows competitive performance on detailed pose estimation and object classification.