Transient LASSO: Transient Large-Scale Scene Reconstruction

Scene Reconstruction and Decomposition

Transient LASSO allows for decomposing measurements into reflectivity and dense geometry using only a single driving trajectory of transient LiDAR data. Our approach is built upon a realistic imaging formulation model that explicitly simulates both the back-reflected and ambient light components in the scene.

Transient Neural Field

We represent the scene as a neural field f : {x} → {σ, n, α, η, A} mapping each point in space x to its volumetric density σ, normal n, albedo α, retroreflectivity η and ambient light A, shown in the figure below. We employ two multi-headed neural fields, to model respectively the geometry and appearance of the scene. Both fields share a spatial embedding χ via multi-resolution hash encoding H.
We then reconstruct the noise-free transient waveform for low- or high-flux conditions, through a time resolved volume rendering approach.

Scene & Sensor Parameter Control

Transient LASSO enables explicit control of scene and sensor parameters. This allows sensor manufacturers and autonomous vehicle developers who can replace expensive sensor tests with Transient LASSO in order to, e.g., determine how certain hardware parameters influence downstream detection performance or point cloud quality in challenging conditions.

To showcase the ability of Transient LASSO for scene and parameter control, we alter the ambient component as well as the laser pulse power.
As visualized below, increasing the ambient light component by a factor 10 (top row), renders the pulse invisible from the background. Conversely, with a higher pulse power, the propagating laser pulse remains visible at greater distances (bottom row).