Desoiling Dataset: Restoring Soiled Areas on Automotive Fisheye Cameras

Surround-view cameras became an integral part of autonomous driving setup. Being directly exposed to harsh environmental settings, they can get soiled easily. When cameras get soiled, the degradation of performance is usually more dramatic compared to other sensors. Having this on mind, we decided to design a dataset for measuring the performance degradation as well as to help constructing classifiers for soiling detection, or for trying to restore the soiled images, so we can increase the performance of the off-the-shelf classifiers. The proposed dataset contains 40+ approximately 1 minute long video sequences with paired image information of both clean and soiled nature. The dataset will be released as a companion to our recently published dataset [14] to encourage further research in this area. We constructed a CycleGAN architecture to produce de-soiled images and demonstrate 5% improvement in road detection and 3% improvement in detection of lanes and curbs.

1. Introduction

The advances in autonomous driving show that combination of sensors is a necessary step to achieve difficult safety and reliability standards. Surround view cameras are becoming de facto standard in autonomous parking, where they significantly contribute to ultrasonic sensors by resolving difficult scenarios, such as fishbone parking or detecting a free parking spot outlined only by ground markings, which is completely not resolvable by using ultrasonic sensors solely. Some influential people even believe that cameras could replace expensive sensors like Lidars.