Two-Edge-Resolved 3d Non-Line-of-Sight Imaging: A Fisher Information Equalized Discretization

Abstract

We demonstrate high-quality 3D non-line-of-sight imaging by exploiting a two-edge occluder and a scene representation constructed through a Fisher information analysis. Our reconstructions are computed from a single 2D photograph, and like many computational imaging problems, this inverse problem is ill-conditioned. The conditioning depends, in part, on the discretization of the hidden scene, both in terms of the coarseness and shape of the discretization grid. We show that a non-uniform discretization constructed to equalize the Fisher information metric improves the conditioning of the inverse problem for a desired reconstruction resolution. Experimental results demonstrate an improvement in condition number of over three orders, enabling more detailed and visually accurate reconstructions without increasing the reconstruction resolution.