Master's thesis

Supervision by: Iro Laina, Christian Rupprecht, Federico Tombari
Thesis by: Marat Seroglazov

Abstract

During the past few years, remarkable progress has been made in the field of semantic image segmentation, thanks to deep learning and CNN advances. However, it remains challenging to reach a high accuracy using training data with image-level annotations only, i.e. in a weakly supervised fashion. In this project, we address the problem of semantic segmentation from image-level labels together with a small percentage of pixel-wise labeled images. Our approach will begin from recent work in estimating class-specific saliency maps [1]. Our goal is to further enhance its training pipeline, introduce a multiple hypothesis prediction approach [2] for improving the quality of the semantic outputs and additionally explore its usability for instance segmentation.

Resources

[1] Shimoda, Wataru, and Keiji Yanai. "Distinct class-specific saliency maps for weakly supervised semantic segmentation." In European Conference on Computer Vision, pp. 218-234. Springer, Cham, 2016.
[2] Rupprecht, Christian, Iro Laina, Robert DiPietro^?, Maximilian Baust, Federico Tombari, Nassir Navab, and Gregory D. Hager. "Learning in an uncertain world: Representing ambiguity through multiple hypotheses." In International Conference on Computer Vision (ICCV). 2017.