Efficient implicit reconstruction of hidden object in two-bounce non-line-of-sight imaging

Recently, Two-bounce NLoS imaging that uses one wall to reflect laser light and one wall to receive the shadows for hidden voxels carving has demonstrated great potential in terms of low-cost off-the-shelf components, where the key is to extract visual features in shadows and use key rays for efficient reconstruction. However, the fineness of hidden target depiction depends on a large number of multi-view rays, which leads to a multiplication of the computational latency. In contrast to the explicit modelling of previous methods, this paper presents a fast ray carving method in an implicit way that uses a hierarchical Neural Radiance Field (NeRF) model to reconstruct the hidden objects. Specifically, the multi-view shadow images are decomposed spatially and temporally to extract important rays that mainly shape the hidden scenes, optimizing the NeRF model that reduces a large number of rays with repeated structures for accurate and efficient hidden object reconstruction. Our method demonstrates efficient hidden object reconstruction capability for both dynamic and static occluded scenes, achieving a relative depth deviation of 2% in a 20m×10m×4m cube, distinguishing fine targets up to 6 cm laterally and 3 cm axially.