Mixed Precision ℓ1 Solver for Compressive Depth Reconstruction: An ADMM Case Study

Abstract

Rapid reconstruction of depth images from sparsely sampled data is important for many machine learning applications, including robot or vehicle assistance or autonomy, which require low power LiDAR sensing for eye safety, and resource reduction for FPGA or solid state implementation, especially with constrained energy budgets. A new compressive depth reconstruction design approach is proposed using a compact ADMM solver for the lasso problem, which varies the precision scaling in an iterative optimization process. Implementations on an FPGA architecture show over 55% savings in hardware resources and 78% in power with only minor reduction in reconstructed depth image quality compared to single float precision.