How good is your algorithm: the significance of data diversity
A. Vranic, Darja Cvetković · 2022
Artificial intelligence (AI) algorithms permeate all spheres of human lives to a great extent. Their influence is expected to only grow in the future. Autonomous driving is no exception and is not possible without AI algorithms. The quality of AI algorithms closely depends on the data used for their training, testing, and validation. Due to the specific nature of how they are collected, the data sets used for autonomous driving algorithms often lack diversity. The majority of the data is collected under ideal weather conditions. The performance of most algorithms during adverse weather conditions is thus unknown. We present an overview of physics-based algorithms used for transforming images and videos taken during ideal weather conditions into ones with haze, fog, and rain. We show that performance algorithms for deep object detection and semantic segmentation algorithms decrease with the increase of weather-induced noise in images and videos. Our survey confirms the need to introduce diverse datasets with well-controlled noise levels for training, testing, and validation of autonomous driving algorithms.