Non-Line-of-Sight (NLOS) imaging uses fast illumination and detection
to reconstruct images of scenes from indirect illumination. Light
reflected off a relay surface is thereby used to view the scene. One
approach to compute these reconstructions from the captured data is to
transform them into line of sight wave propagation problems creating a
virtual wave front to model a virtual camera at the location of the
relay surface. Our NLOS imaging system samples the virtual wavefront
at the virtual aperture. As in a line of sight camera, scene
reconstruction for the virtual camera is achieved by propagating the
virtual wave back into the scene. In line of sight cameras, this
operation is often performed by a lens. For the virtual camera, we
implement it computationally. This approach allows us to transfer
methods for scene reconstruction, scene inference, and imaging from
existing line of sight imaging approaches to NLOS imaging.
In particular, we make use of fast wave propagation algorithms to
create high speed memory efficient NLOS imaging. This allows us to
reconstruct complex scenes in sub-second times for variable hardware
configurations. In particular, our reconstruction methods allow us to
use SPAD arrays in conjunction with laser scanning to improve capture
speed.
There is a large diversity of line of sight imaging approaches with
different properties that probe different aspects of the scene. In
principle, the Phasor Field Virtual Wave formalism allows us to turn
any of them into a NLOS virtual camera. I will cover several examples
of this process that yield different NLOS reconstructions, including
2D NLOS imaging, transient NLOS videos, and Visualization of higher
order light paths from 4th and 5th bounces in the hidden scene.
|