Non-line-of-sight imaging with adaptive artifact cancellation

Abstract

Non-Line-of-Sight (NLOS) image reconstruction algorithms commonly encounter a significant challenge: their dependence on scenario and empirically derived parameters, which undermines the algorithms’ generalizability and adaptiveness. To tackle this problem, we have devised a forward projection model and a novel evaluation metric for NLOS reconstruction, named Time-of-Flight Structural Similarity (TOF-SSIM). This metric is independent of ground truth and serves to assess image quality and systematically determine the optimal parameters of reconstruction algorithms. Within this method, this paper presents an Adaptive Artifact Cancellation (AAC) algorithm. We first generate a sequence of new Time-of-Flight (TOF) histograms by subtracting the TOF histogram convolved with a Gaussian kernel from the original TOF histogram. Subsequently, the new TOF histograms are backprojected to reconstruct an artifact-reduced image. Our method has been validated using both our datasets and public datasets, evidencing that the AAC algorithm excels in producing efficient and consistent reconstructions under both confocal and non-confocal configurations, with reconstruction quality comparable to other state-of-the-art algorithms.