Increasing the Diversity in RGB-to-Thermal Image Translation for Automotive Applications

Abstract

Thermal imaging technology in Advanced Driver Assistance Systems (ADAS) enhances road safety through its superior perception capabilities in low-light and harsh-weather conditions, as compared to traditional RGB cameras. However, research into this technology and its associated artificial intelligence (AI) models is hampered by inferior dataset availability and lack of representation in driving simulators. RGB-to-thermal image translation offers a promising solution. Whereas existing methods focus on learning a one-to-one RGB-to-thermal mapping, we argue that the mapping ought to be one-to-many. To achieve this, we leverage a multi-modal translation framework and equip it with our Component-aware Adaptive Instance Normalization (CoAdaIN), introduced in this work. In contrast to the original AdaIN, which operates globally, CoAdaIN enables the application of various styles to different image components individually. The result, as we show, is more realistic and diverse thermal image translations.