Monocular Depth Estimation from Thermal Images Via Edge-Guided Rgb Synthesis

Monocular depth estimation from thermal (infrared, IR) images remains underexplored, mainly due to the lack of datasets and the high cost of training large models. Standard models are trained on daylight RGB images and do not perform as well on thermal images. We propose an alternative method: Generate realistic RGB images from IR images using state-of-the-art generative models with edge control. And use these generated images with the most advanced monocular depth estimation models to get more accurate depth maps. We evaluated our method on KAIST [1] multispectral RGB-IR dataset, using RGB images to obtain ground-truth depth maps. We achieved significant improvements in RMSE, AbsRel and Delta scores.