Statistical Part-based Models for Object Detection in Large 3D Scans

Abstract

3D scanning technology has matured to a point where very large scale acquisition of high resolution geometry has become feasible. However, having large quantities of 3D data poses new technical challenges. Many applications of practical use require an understanding of semantics of the acquired geometry. Consequently scene understanding plays a key role for many applications.This thesis is concerned with two core topics: 3D object detection and semantic alignment. We address the problem of efficiently detecting large quantities of objects in 3D scans according to object categories learned from sparse user annotation. Objects are modeled by a collection of smaller sub-parts and a graph structure representing part dependencies. The thesis introduces two novel approaches: A part-based chain structured Markov model and a general part-based full correlation model. Both models come with efficient detection schemes which allow for interactive run-times.