Robust Semantic Video Segmentation through Confidence-based Feature Map Warping
Timo Sämann, Karl Amende, Stefan Milz, Horst–Michael Groß · 2019
One of the limiting factors when using deep learning methods in the field of highly automated driving is their lack of robustness. Objects that suddenly appear or disappear from one image to another due to inaccurate predictions as well as occurring perturbations in the input data can have devastating consequences. A possibility to increase model robustness is the use of temporal consistency in video data. Our approach aims for a confidence-based combination of feature maps that are warped from previous time stages into the current one. This enables us to stabilize the network prediction and increase its robustness against perturbations. In order to demonstrate the effectiveness of our approach, we have created a test data set with image perturbations such as image artifacts and adversarial examples in which we significantly outperform the baseline.