Dear all,
you are cordially invited to join the project presentation of Simon Kuhn on 07.10.2021 at 16:15. You can join via Zoom as indicated in the course details.
Topic: Verifying Convolutional Neural Networks: A Landmark-based Approach for Quantitative Evaluation of Facial Action Unit Classication
Abstract:
The transparency of neural networks' decisions is crucial in certain areas such as human facial pain detection. In order to automatically evaluate whether a neural net identified class-relevant parts of faces, a verification pipeline was introduced previously. The pipeline uses layerwise relevance propagation and bounding boxes around class-relevant facial areas for evaluation. We extended this approach using polygons as refined boundary instead of boxes. These polygons are defined in such way that they yield a more precise approximation of the class-relevant facial areas compared to the bounding boxes. This minimizes the risk of class-irrelevant facial areas and background contributing to higher evaluation scores. We compare our polygonal boundary approach to the previous bounding box approach. Our results show the necessity of refined boundaries in order to eliminate false relevance attributions.