Sites Inria

Version française

Junior Seminar

Junior Seminar - 3D Scene Understanding from Images

3D Scene Understanding from Images

Vincent Lepetit, Université de Bordeaux

3D scene understanding is an old, yet unsolved, fundamental Computer Vision problem, which aims at recognising objects and estimating accurately their spatial locations, orientations, dynamics, shapes, and interactions from simple images or videos streams.  It is a very challenging task, because lighting, perspective projection, and partial occlusions make images very difficult to analyze, especially for 3D perception.  Moreover, the "understanding task" can be ambiguous as its desired output may vary with the final application.
In this talk, I will present the research developed in my group on 3D scene understanding. I will mainly focus on two recent works. The first work introduces a general method for 3D pose estimation based on a feedback loop inspired by biological mechanisms, and providing very accurate and reliable results. The second work shows how to modify standard Deep Networks to make them resilient to partial occlusions. It is also very general and can be applied to different problems.