Sebastian Nowozin
Wednesday 21st October 2015
Time: 4.00pm
Ground Floor Seminar Room
25 Howland Street, London, W1t 4JG
Bayesian Time-of-Flight for Realtime Shape, Illumination and Albedo
Humans have a remarkable ability to understand their environment through their
visual sense but in recent years computers have made rapid progress and
approach or exceed human-level performance on some tasks such as optical
inspection, motion tracking, and image classification.
One recent technology that enabled much of this progress are depth cameras
which reconstruct in real-time the scene geometry independent from light and
texture, thus simplifying high level tasks such as recognition and
3D reconstruction.
In this talk, I will give an accessible overview to non-experts about a fully
Bayesian model behind a recent state-of-the-art depth sensing system. This
system is based on the principle of time-of-flight, that is, timing light
pulses as they are reflected of surfaces. This sensing modality is not used
by any animal and requires interesting design considerations which we address
using a principled decision-theoretic framework. Furthermore significant
challenges exist in real-time implementation and in addressing systematic
model violations due to multiple light bounce phenomena.
Joint work with Amit Adam, Christoph Dann, Omer Yair, and Shai Mazor.
(A background in probabilistic models and machine learning will be sufficient
to understand the talk.)