Object recognition and scene parsing using 3D cues

Martial Hebert
Carnegie-Mellon University
The Robotics Institute

3D geometric cues are crucial to understanding scenes. For example, constraints on the relative placement of object and surfaces in a scene can be properly expressed only in 3D space. However, the information necessary for 3D geometric reasoning is not directly available from a single input image. In this talk, we will discuss different ways to incorporate 3D reasoning in image interpretation. In particular, we will discuss recent techniques to incorporate 3D constraints such as interposition and occlusion, volumetric and physical constraints, relative placement of objects, and functional constraints on 3D scene configurations. We will also briefly discuss non-parametric, data-driven approaches to 3D reasoning from images.

Presentation (PDF File)

Back to Graduate Summer School: Computer Vision