Effect of image quality in computer vision for semantic segmentation of road images

Convolutional neural networks have real practical application potential, such as autonomous driving, but they are known to be sensitive to image degradation. The focus of this research is to give insight into the robustness of the current state-of-the-art model for semantic segmentation against corruptions likely to be encountered in real settings, specifically compression, motion blur and Poisson (shot) noise. In a safety-critical application, the precise semantic segmentation of certain instance classes, for example persons or vehicles, can be considered more important than others, such as vegetation or the sky, which is why the robustness of individual instance classes is also assessed with the intent to determine model deployability.