Podrobnosti studentského projektu

Téma:Universal visual representation with deep learning
Katedra:Katedra kybernetiky
Vedoucí:Georgios Tolias, Ph.D.
Vypsáno jako:Diplomová práce, Bakalářská práce, Semestrální projekt
Popis:Training a convolutional neural network to generate descriptors is typically performed on training sets coming from the domain of the target application. For example, training is performed on datasets of landmarks, or logos, or retail products, for landmark, or logo, or retail product recognition respectively. Each trained model performs well on the corresponding domain and worse on the others. The aim of this project is to study the generalization properties of such different models, and the overlap between different domains. Then, a joint model will be trained to handle all tasks in a universal way. The training will be performed in a unified dataset, while balancing the focus on different domains and on different visual cues sush as shapes and textures.
Literatura:Radenovic Tolias Chum, PAMI 2019, Fine-tuning CNN Image Retrieval with No Human Annotation
Vo Hays, WACV 2018, Generalization in Metric Learning: Should the Embedding Layer be the Embedding Layer
Musgrave, Belongie, Lim, ECCV 2020, A Metric Learning Reality Check
Za obsah zodpovídá: Petr Pošík