Towards Robust CNN-Based Object Detection through Augmentation with Synthetic Rain Variations

Abstract

Convolutional Neural Networks (CNNs) achieve high accuracy in vision-based object detection tasks. For their usage in the automotive domain, CNNs have to be robust against various kinds of natural distortions caused by different weather conditions while state-of-the-art datasets like KITTI lack these challenging scenarios. Our approach automatically identifies corner cases where CNNs fail and improves their robustness by automated augmentation of the training data with synthetic rain variations including falling rain with brightness reduction as well as raindrops on the windshield. Our method achieves higher performance upon validation against a real rain dataset compared with state-of-the art data augmentation techniques like Gaussian noise (GN) or Salt-and-Pepper noise (SPN).