Simulating weather conditions on digital images

Abstract

Object detection is an essential part of Autonomous Driving and Driver-Assistance systems. Nevertheless, bad weather conditions like fog, rain and smog decrease the ability to understand the surrounding scene, just like human vision can be impaired in such situations. To overcome this challenge, we simulate fog on clear images that are already annotated then systems are trained to recognize objects using such dataset. In this thesis, we use Convolutional Neural Networks to generate fog on clear images, reducing the human effort of finding mathematical foggificaion formulas and leting our model figure that out. We propose a generative model that simulates fog on clear images using an unpaired dataset of both foggy and fog-free images.