Augmenting Image Data Sets With Water Spray Caused by Vehicles on Wet Roads

Adverse weather conditions challenge object detection neural networks, because they are mostly trained on clean data sets that were taken in good weather. But autonomous vehicles rely on accurate detection and classification of other road users for safe and reliable operation. Capturing new data sets in rough weather is time consuming and expensive. We therefore propose a method to augment existing image data sets with physically realistic water spray swirled up by vehicles—one of the most influential disturbances for even human drivers on wet roads in or after heavy rain. Using a wide range of newly augmented images, we evaluate the influence on an established convolutional neural network object detection and showcase the potential of training with these new augmented data sets.