Automatic Labeling for Thermal Imaging Datasets Generation
Daniel Cantón, María T. Lázaro · 2024
Low visibility scenarios present a significant challenge in the development of robust perception systems due to the difficulty in replicating real conditions and the lack of available data for training neural networks across a wide range of situations. This problem is even more pronounced when using less common sensors such as thermal (infrared) cameras, which operate in the non-visible spectrum and provide better performance in low light, dusty, or smoky conditions. This paper addresses the problem of dataset generation in such scenarios by proposing an automatic labeling method that leverages the capabilities of pre-trained networks on images captured in the visible spectrum using traditional cameras. We evaluate the proposed method by comparing the quality of automatically generated datasets with manually annotated datasets. Finally, we demonstrate the versatility of training a network in the non-visible spectrum and applying it to low visibility situations.