Connectionist Models for Learning Local Image Descriptors: An empirical case study

Christian Osendorfer

Abstract: This thesis demonstrates that supervised as well as unsupervised Neural Network-based approaches can learn compact descriptors for local image patches. Additionally it is shown that Explicit Negative Contrasting improves multi-view graphical models. Also, Hobbesian Networks are introduced, utilizing differential equations to induce deep models. Finally, vaeRIM combines variational inference with unsupervised clustering in a novel way.