Parallax-Driven Denoising of Passive Non-Line-of-Sight Thermal Imagery

Connor Hashemi, Takahiro Sasaki, James Leger

Abstract: Current thermal non-line-of-sight (NLOS) imaging is most effective when applied to hidden objects with substantial thermal contrast reflecting off homogeneous scattering surfaces. However, this ideal scenario is not realistic in many practical situations. Incorporating the light field, which includes capturing multiple pictures at different camera positions, can provide additional information that is useful in denoising weak signals. This paper introduces a two-part algorithm that utilizes geometrical parallax from hidden thermal objects to improve reconstruction of objects with temperatures only slightly above ambient and where the scattering surface is highly inhomogeneous. The first algorithm uses both image and light field coordinate frames to remove spatial sensor noise and the self-radiant component of the scattering surface. The second uses parallax information contained in the 4-D light field to denoise the scattered radiance and estimate the 3-D location of the hidden objects. We demonstrate our methods using long-wave infrared radiation from a human subject scattered off a spatially-varying surface and achieve improved contrast-to-noise and normalized correlation metrics compared to another popular denoising technique.