ALOP Colloquium with Hanno Gottschalk

On Tuesday, March 3 2020 at 16:00 c.t.  Prof. Dr. Hanno Gottschalk  , Bergische Universität Wuppertal  will present his recent work. entitled

False Positives and False Negatives in Semantic Segmentation: Toward the assurance of the perception with deep CNN in atonomous driving

Abstract:

The application of machine learning technology in safety relevant environments poses several challenges regarding their reliability. This especially applies to modern deep learning architectures, which pretty much resemble a black box. In this talk, measures of uncertainty from basic classification algorithms to current fully convolutional networks for semantic segmentation are discussed. We present known and new uncertainty measures and use them in meta classification: An algorithmic decision, if a prediction by the AI is correct of not, without access to the ground truth data. For the case of semantic segmentation, we use this for the instance based detection of false positives. Techniques for the detection of false negatives are discussed as well. Both techniques can be fused for a more reliable detection of underrepresented classes, like pedestrians in street scenes.

The presentation will take place in HS 9.

Please join us for coffee at 15:45 in E10.