[Uk Cheol Shin, Kwanyong Park and Byeong-Uk Lee and professor In So Kweon from left]
WACV is a major academic conference ranked 9th in terms of the Google Scholar h-5 index within the field of computer vision.
The KAIST team’s award-winning work was among this year’s 641 published papers, which were by themselves selected out of 1,577 submission.
Titled ” Self-supervised Monocular Depth Estimation from Thermal Images via Adversarial Multi-spectral Adaptation”, their paper deals with the estimation of distance from a single thermal image, as one of the most difficult problems in computer vision involving challenges with the low resolution of thermal images and the lack of detailed image data labelled with temperature distribution.
To address this problem, the team proposed a novel deep learning model that combines self-supervised learning with adversarial learning between multispectral images.
Unlike conventional methods that are limited by constraints such as the requirement of the exact camera settings, the model is able to learn without these constraints through the utilization of individual thermal and color images.
The model was tested in various experimental conditions such as day, night and poor illumination, and high performance was achieved under various conditions compared to the existing method.