Joint self-supervised learning and adversarial adaptation for monocular depth estimation from thermal image

Depth estimation from thermal images is one potential solution to achieve reliability and robustness against diverse weather, lighting, and environmental conditions. Also, a self-supervised training method further boosts its scalability to various scenarios, which are usually impossible to collect ground-truth labels, such as GPS-denied and LiDAR-denied conditions. However, self-supervision from thermal images is usually insufficient to train networks due to the thermal image properties, such as low-contrast and textureless properties. Introducing additional self-supervision sources (e.g., RGB images) also introduces further hardware and software constraints, such as complicated multi-sensor calibration and synchronized data acquisition. Therefore, this manuscript proposes a novel training framework combining self-supervised learning and adversarial feature adaptation to leverage additional modality information without such constraints. The framework aims to train a network that estimates a monocular depth map from a thermal image in a self-supervised manner. In the training stage, the framework utilizes two self-supervisions; image reconstruction of unpaired RGB-thermal images and adversarial feature adaptation between unpaired RGB-thermal features. Based on the proposed method, the trained network achieves state-of-the-art quantitative results and edge-preserved depth estimation results compared to previous methods. Our source code is available at www.github.com/ukcheolshin/SelfDepth4Thermal