Exploring the Usage of Diffusion Models for Thermal Image Super-resolution: A Generic Uncertainty-aware Approach for Guided and Non-guided Schemes

Carlos Cortés-Mendez, Jean-Bernard Hayet; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2024, pp. 3123-3130

In this paper we explore the use of diffusion models for the thermal imaging super-resolution problem with the PBVS workshop Thermal Image Super-Resolution Challenge (TISR) as an application context. In addition of adapting the recently proposed ResShift diffusion approach to the problem of SR for thermal imaging we show how this diffusion model can be also used nearly effortless in both the guided and non-guided TISR tasks where the guidance comes from the visible image. More crucially we show that a natural and often under-leveraged output from this diffusion approach is the quantification of the aleatoric uncertainty on the resulting HR prediction. By using this property we empirically show that per-pixel standard deviation of the samples produced by a super-resolution diffusion model are a good estimator for the per-pixel absolute error in scenarios where the HR ground truth is not available.