Detail of the student project

Topic:Dense visual recognition models trained with sparse annotations: application to drought watch from satellite images
Department:Katedra kybernetiky
Supervisor:Georgios Tolias, Ph.D.
Announce as:Diplomová práce, Bakalářská práce, Semestrální projekt
Description:Semantic segmentation is the task of predicting the semantic class of each image pixel. It typically involves dense prediction models that are trained with images that are annotated densely, i.e. each pixel is annotated. Sparse annotation is much easier to obtain, for instance when labeling only a few pixels per training image. The goal of this project is to use sparse annotation to train a dense prediction model for semantic segmentation. One of the target applications is drought detection through satellite images. The training data for drought conditions are provided by human experts in Northern Kenya in a sparse way (one pixel is annotated) through ground-level images. In this task, dense annotation for training is not available at all. The project can focus on other applications of semantic segmentation as well, e.g. semantic segmentation in autonomous driving, building facade segmentation, semantic segmentation of satellite images, etc.
Bibliography:A Bearman, O Russakovsky, V Ferrari, L Fei-Fei, ECCV 2016, What’s the point: Semantic segmentation with point supervision
Vernaza Chandraker, CVPR 2017, Learning random-walk label propagation for weakly-supervised semantic segmentation
Responsible person: Petr Pošík