IC Colloquium: Uncertainty Quantification and Label Error Detection for Semantic Segmentation
By: Matthias Rottmann - University of Wuppertal
Video of his talk
Abstract
Currently, deep learning is one of the most powerful tools for the automation of complex tasks, such as environmental perception in automated driving or robotics. However, the lack of proper assessment of the reliability of a given prediction often hinders the successful application of deep learning in practice. In this talk, we consider the semantic segmentation of camera images in the context of street scenes. I introduce an uncertainty quantification method for predictions of deep neural networks on the level of predicted connected components, show performance results and its application to the detection of label errors in semantic segmentation datasets. Besides that, we get a glimpse of further applications.
Bio
Matthias Rottmann achieved his PhD in applied mathematics in 2016 at University of Wuppertal (UW), Germany. He was postdoc at UW's applied computer science group from 2016 to 2020. In 2020 he got assigned a tenured lecturer position at UW's stochastics group. He is leading a junior research group with focus on uncertainty in deep learning and is currently visiting researcher at EPFL.
Video of his talk
Abstract
Currently, deep learning is one of the most powerful tools for the automation of complex tasks, such as environmental perception in automated driving or robotics. However, the lack of proper assessment of the reliability of a given prediction often hinders the successful application of deep learning in practice. In this talk, we consider the semantic segmentation of camera images in the context of street scenes. I introduce an uncertainty quantification method for predictions of deep neural networks on the level of predicted connected components, show performance results and its application to the detection of label errors in semantic segmentation datasets. Besides that, we get a glimpse of further applications.
Bio
Matthias Rottmann achieved his PhD in applied mathematics in 2016 at University of Wuppertal (UW), Germany. He was postdoc at UW's applied computer science group from 2016 to 2020. In 2020 he got assigned a tenured lecturer position at UW's stochastics group. He is leading a junior research group with focus on uncertainty in deep learning and is currently visiting researcher at EPFL.
Practical information
- General public
- Free
- This event is internal
Contact
- Host: Mathieu Salzmann