Department of Computer Engineering
CS 590 SEMINAR
Object-based Unsupervised Image Localization and Segmentation with Recurrent Attentive Networks
Yarkın Deniz Çetin
Computer Engineering Department
Image segmentation and object localization have many real-world applications. Current localization methods depend on using supervised training which requires vast amounts of labeled data. These data can be difficult obtain in sufficient sizes for particular domains such as biology. We instead propose a recurrent attention-based model which exploits the co-occurrence of pixels in objects to localize and extract object locations without supervision and without pre-trained models. Our model is an extension of deep recurrent attentive writer (DRAW) unsupervised image localization model. We train our model by reconstructing given scenes. Using a soft attention-based one-object-at-a-time network we are able to learn a model which can localize individual objects in the scene.
DATE: 12 November, 2018, Monday, CS590 presentations begin at @ 15:40