Sharing matters for generalization in deep metric learning

• Verfasst von: Milbich, Timo [VerfasserIn]
• Verfasst von: Roth, Karsten [VerfasserIn]
• Verfasst von: Brattoli, Biagio [VerfasserIn]
• Verfasst von: Ommer, Björn [VerfasserIn]
• Titel: Sharing matters for generalization in deep metric learning
• Verf.angabe: Timo Milbich, Karsten Roth, Biagio Brattoli, and Björn Ommer
• Jahr: 2022
• Umfang: 12 S.
• Fussnoten: Gesehen am 04.01.2022 ; Date of Publication: 15 July 2020 ; Date of current version 3 Dec. 2021
• Titel Quelle: Enthalten in: Institute of Electrical and Electronics EngineersIEEE transactions on pattern analysis and machine intelligence
• Ort Quelle: New York, NY : IEEE, 1979
• Jahr Quelle: 2022
• Band/Heft Quelle: 44(2022), 1, Seite 416-427
• ISSN Quelle: 1939-3539
• DOI: doi:10.1109/TPAMI.2020.3009620
• Datenträger: Online-Ressource
• Sprache: eng
• Sach-SW: deep learning
• Sach-SW: Deep metric learning
• Sach-SW: Encoding
• Sach-SW: generalization
• Sach-SW: Image color analysis
• Sach-SW: image retrieval
• Sach-SW: Measurement
• Sach-SW: shared features
• Sach-SW: similarity learning
• Sach-SW: Standards
• Sach-SW: Task analysis
• Sach-SW: Training
• Sach-SW: Training data
• K10plus-PPN: 1784559024

Abstract

Learning the similarity between images constitutes the foundation for numerous vision tasks. The common paradigm is discriminative metric learning, which seeks an embedding that separates different training classes. However, the main challenge is to learn a metric that not only generalizes from training to novel, but related, test samples. It should also transfer to different object classes. So what complementary information is missed by the discriminative paradigm? Besides finding characteristics that separate between classes, we also need them to likely occur in novel categories, which is indicated if they are shared across training classes. This work investigates how to learn such characteristics without the need for extra annotations or training data. By formulating our approach as a novel triplet sampling strategy, it can be easily applied on top of recent ranking loss frameworks. Experiments show that, independent of the underlying network architecture and the specific ranking loss, our approach significantly improves performance in deep metric learning, leading to new the state-of-the-art results on various standard benchmark datasets.