Computationally efficient pixelwise deep learning architecture for accurate depth reconstruction for single-photon LiDAR

Abstract

This work introduces a compact deep learning architecture for depth image reconstruction from time-resolved single-photon histograms. Unlike most deep learning approaches that mainly rely on 3D convolutions, our network is implemented purely with 1D convolutions without assistance from other sensors or pre-processing. Both synthetic and real datasets were used to evaluate the accuracy of our model for challenging signal-to-background ratios (SBRs), ranging from 5:1 to 1:1. Conventional maximum likelihood (ML) and another photon-efficient optimization-based algorithm were adopted for performance comparisons. Results from synthetic data show that our model achieves lower mean absolute error (MAE). Additionally, results from real data indicate that our model exhibits better reconstruction for high-ambient effects and provides better spatial information. Unlike existing 3D deep learning models, we process pixel-wise histograms continuously, rather than splitting the point cloud and stitching them afterward, which saves memory and computational resources, thereby laying a foundation for real-world embedded applications.