Embedded Systems

Simulating Photo-Realistic Snow and Fog on Existing Images for Enhanced CNN Training and Evaluation

by Alexander von Bernuth, Georg Volk, and Oliver Bringmann
In 2019 IEEE Intelligent Transportation Systems Conference (ITSC), pages 41-46, 2019.

Abstract

Verification and robustness testing of machine learning algorithms for autonomous driving is crucial. Due to the increasing complexity and quantity of those systems in a single vehicle, just driving the required distance with a newly developed vehicle is not feasible anymore: billions of hours on the street without failure are necessary to qualify for industry standards like ISO 26262. That is where simulation comes into play: machine learning algorithms are trained and evaluated on well known image data sets like KITTI or Cityscapes. But today’s data sets mostly contain images taken under perfect weather conditions and therefore do not harden optical object detection algorithms against various weather conditions. This paper focuses on reusing these established and labeled data sets by augmenting them with adverse weather effects like snow and fog. Those effects are rendered physically correct and life like while being added to existing real world images. Thanks to easy parametrization the weather influences may be varied as necessary and allow for finely tuned learning and optimization processes. The weather effects are evaluated with regard to realism and impact on an established object detection algorithm. These newly created weather-influenced images may be used to validate or train new object detection algorithms.