Simulating Photo-realistic Snow and Fog on Existing Images for Enhanced CNN Training and Evaluation
Authors:
- 1 October 2019
- conference paper
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 41-46
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.Keywords
This publication has 22 references indexed in Scilit:
- Quo vadis, FAS?2015
- Real-Times Snowfall Simulation Based on Particle System and PulverizationApplied Mechanics and Materials, 2013
- Visual 3D Perception of Motion Environment and Visibility Factors in Virtual Space2012
- An innovative artificial fog production device improved in the European project “FOG”Atmospheric Research, 2007
- Real-time snowing simulationThe Visual Computer, 2006
- Rendering falling rain and snow2004
- Visual simulation of snowfall, snow cover and snowmelt2002
- Semi-Monte Carlo Light Tracing Applied to the Study of Road Visibility in Fog2000
- Snow mass concentration and precipitation rateCold Regions Science and Technology, 1988
- Modeling the effect of the atmosphere on lightACM Transactions on Graphics, 1987