Point- and volume-based neural rendering

Neural rendering has recently emerged as a powerful paradigm for creating novel views from a set of input images. Broadly speaking, neural rendering methods rely on a set of photographs for learning a representation of the underlying scene. The learned representation can subsequently be queried to generate images of that same scene from a viewpoint that was not previously captured. In this internship, we will explore novel methods for combining existing neural rendering techniques, namely: point- and volume-based. These two techniques are complementary to each other: volume-based methods are flexible and can learn complex geometry, while point-based techniques are fast. If successful, this research will lead to advances which will likely have impact in virtual and augmented reality, historical preservation, and entertainment.

Faculty Supervisor:

Jean-Francois Lalonde

Student:

Partner:

Inria Sophia Antipolis - Méditerranée Research Centre

Discipline:

Computer science

Sector:

Education

University:

Université Laval

Program: