Exploiting & Refining Depth Distributions With Triangulation Light Curtains

RMSE: 1.2m RMSE: 0.4m (a) (b) (c) (d) Figure 1: (a) We show how errors in Monocular Depth Estimation are corrected when used in tandem with an Adaptive Sensor such as a Triangulating Light Curtain (Yellow Points and Red lines are Ground Truth). (b) We predict a per-pixel Depth Probability Volume from Monocular RGB and we observe large per-pixel uncertainties (σ = 3m) as seen in the Bird's Eye View / Top-Down Uncertainty Field slice. (c) We actively drive the Light Curtain sensor's Laser to exploit and sense multiple regions along a curve that maximize information gained. (d) We feed these measurements back recursively to get a refined depth estimate, along with a reduction in uncertainty (σ = 1m).