Duminil, Alexandra (Université Gustave Eiffel), Ieng, Sio-Song (Université Gustave Eiffel), Gruyer, Dominique (Université Gustave Eiffel)

Proposal of Fidelity Scores for Synthetic Image Evaluation and Validation in Degraded Weather Conditions

Scheduled for presentation during the Regular Session "Synthetic datasets in perception" (ThBT4), Thursday, September 26, 2024, 14:50−15:10, Salon 7

2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC), September 24- 27, 2024, Edmonton, Canada

This information is tentative and subject to change. Compiled on December 26, 2024

Abstract

The generation of large quantities of realistic road scenes and scenarios that include adverse weather conditions is essential in the field of robust advanced driving systems, especially for both training deep learning-based methods and validation processes. In order to generate such a large datasets, lots of works have been done with the objective of generating increasingly large and realistic synthetic datasets using graphics engines or synthetic-to-real domain adaptation algorithms. Our research aims to adopt a comprehensive and generic conceptual framework to quantify the level of fidelity of the computed-generated RGB images in degraded weather conditions, unlike existing methods that are predominantly application-specific. To reach this goal, a set of scores based on different features for assessing the level of fidelity of virtual RGB images is proposed. The scores provide probability measurements and reflect in what aspect the computed-generated data has high or low fidelity relatively to real data coming from real cameras. Specifically, we delve into the analysis of texture and high-frequency information in images, specifically focusing on the statistical characteristics of a large set of realistic and synthetic road datasets involving clear, rainy, and foggy conditions. In this paper, a first proof of concept is presented and gives an initial validation of this innovative approach in an urban environment under adverse weather conditions. The obtained results demonstrate the relevance and effectiveness of our method.