List |
Topic: | Use video consistency as supervision for deep learning |
---|---|
Department: | Katedra kybernetiky |
Supervisor: | Georgios Tolias, Ph.D. |
Announce as: | Diplomová práce, Bakalářská práce, Semestrální projekt |
Description: | Training deep convolutional neural networks for visual recognition is a training data demanding task. Human annotators need to manually label very large image collections in order to obtain top performing models. This project aims to use a smaller set of manually labeled examples, but exploit the temporal dimension in videos to enrich the training set. Label propagation models will exploit temporal consistency to obtain confident pseudo-label or temporal inconsistency to pick samples that need further manual annotation. The project will cover both classical label propagation approaches and modern ones that are based on graph convolutional networks. |
Bibliography: | Iscen, Tolias, Avrithis, Chum, CVPR 2019, Label Propagation for Deep Semi-supervised Learning
Zhou, Bousquet, Lal, Weston, Scholkopf, NeurIPS 2004, Learning with Local and Global Consistency Kipf Welling, ICLR 2017, SEMI-SUPERVISED CLASSIFICATION WITH GRAPH CONVOLUTIONAL NETWORKS |