Thermal Non-Line-of-Sight Imaging through Rough Surfaces

Thermal non-line-of-sight (NLOS) imaging utilizes the thermal radiation emitted by hidden objects to “see around corners”, offering promising applications in security and search-and-rescue missions. However, reconstructing hidden scenes through rough relay walls remains highly challenging due to severe spatial mixing caused by complex reflection and scattering. While existing model-based and learning-based approaches yield high-quality results under reflective surfaces, they struggle to deliver satisfactory performance under rough surfaces. Here we introduce NLOSFormer, a physics-embedded neural network to address this challenge. By formulating light transport as a convolution process, NLOSFormer estimates the convolution kernel from measurements to guide reconstruction within an end-to-end network. This explicitly estimated kernel effectively narrows the solution space, significantly improving generalization and enabling relative depth estimation. To support training and evaluation, we further present a publicly available dataset named ThermalNLOS, together with a novel data augmentation strategy to alleviate overfitting. Extensive experiments demonstrate that NLOSFormer significantly outperforms existing approaches across a wide range of relay walls. Notably, it achieves real-time imaging of dynamic targets under rough surfaces, representing a substantial advance toward robust and practical deployment.