DefensePeter

Chair for Computer Aided Medical Procedures & Augmented Reality
Lehrstuhl für Informatikanwendungen in der Medizin & Augmented Reality

THIS WEBPAGE IS DEPRECATED - please visit our new website

PhD Defense by Loïc Peter


  • Speaker: Loïc Peter
  • Date: Friday, September 8th, 2017
  • Time: 10:30 PM
  • Location: FMI-Building, Room 00.12.019

Automating and Assisting Image Segmentation with Decision Forests

Abstract:

The task of semantic image segmentation, i.e. the simultaneous partitioning of an image into regions and the labeling of their content, is a fundamental objective in computer vision and medical image analysis. The need for segmentation algorithms can be encountered in diverse real-world scenarios. In some cases, the segmentation of a scene must be entirely automated, for example to act as a scene understanding tool to help the decision-making of an autonomous system. In other situations, a segmentation algorithm can assist a human user in annotating an image. In this case, the fully-automated aspect is no longer crucial and can be relaxed in favor of an improved flexibility via interactions with the user.

In this thesis, we demonstrate how the statistical learning framework of decision forests can be exploited for these two scenarios and propose three contributions. First, we introduce an efficient and generic scale-adaptive forest training scheme for fully automated image understanding. In the context of large digital slides in histopathology, we propose an interactive and adaptive solution entangling segmentation with image exploration to facilitate the examination of these large images. Finally, we show how the segmentation task can be modeled as a Twenty Questions game towards hands-free human-computer interactions. Overall, we hope to demonstrate the effectiveness of decision forests for automated segmentation in diverse application domains. Moreover, we wish to emphasize the relevance of interactive settings and to introduce practical learning-based solutions in cases where standard interactive techniques are not directly applicable.


WebEventForm
Title: PhD? Defense by Loïc Peter
Date: 8 September 2017
Location:  
Abstract: The task of semantic image segmentation, i.e. the simultaneous partitioning of an image into regions and the labeling of their content, is a fundamental objective in computer vision and medical image analysis. The need for segmentation algorithms can be encountered in diverse real-world scenarios. In some cases, the segmentation of a scene must be entirely automated, for example to act as a scene understanding tool to help the decision-making of an autonomous system. In other situations, a segmentation algorithm can assist a human user in annotating an image. In this case, the fully-automated aspect is no longer crucial and can be relaxed in favor of an improved flexibility via interactions with the user. In this thesis, we demonstrate how the statistical learning framework of decision forests can be exploited for these two scenarios and propose three contributions. First, we introduce an efficient and generic scale-adaptive forest training scheme for fully automated image understanding. In the context of large digital slides in histopathology, we propose an interactive and adaptive solution entangling segmentation with image exploration to facilitate the examination of these large images. Finally, we show how the segmentation task can be modeled as a Twenty Questions game towards hands-free human-computer interactions. Overall, we hope to demonstrate the effectiveness of decision forests for automated segmentation in diverse application domains. Moreover, we wish to emphasize the relevance of interactive settings and to introduce practical learning-based solutions in cases where standard interactive techniques are not directly applicable.
Imageurl: gradhat_small.jpg
Type: News
Videourl:  
Conferencelink:  


Edit | Attach | Refresh | Diffs | More | Revision r1.1 - 07 Sep 2017 - 20:01 - ShadiAlbarqouni

Lehrstuhl für Computer Aided Medical Procedures & Augmented Reality    rss.gif