École Nationale Supérieure de
Techniques Avancées

Résumé

In this document, we remind the reader of the state of the art in the field of single image 3D scene understanding, and in particular of recent work involving the identification of 3D­primitives. We then introduce a new algorithm that makes use of those 3D -­‐primitives and of a few heuristics to create consistent hypotheses on the geometric configuration of a scene and to choose the one most likely to describe the scene. Results are encouraging, but the quality of the results depends heavily on the dataset use d to learn 3D -­‐primitives an how to choose the best hypothesis.