Quality, Efficiency and Reliability of 3D image surface reconstruction
Dragan Tubic
Patrick Hébert (Supervisor)
Denis Laurendeau (Co-supervisor)
Problem: Traditionally, 3D modeling, which consists in creating a virtual 3D model from surface measurements and real objects, is composed of the following three steps which are applied sequentially: acquisition, registration and integration of multiple views. Due to the complexity of existing computation methods, the application of these three steps does not allow the creation of interactive systems where the reconstructed model is available during acquisition. The objective of this project is to unify all three modeling steps and to allow the reconstruction of models irrespective of the type of range data. This project is specifically aimed at hand-held sensors and real time interactive acquisition systems.
Motivation: The acquisition of range data and the reconstruction of models are usually conducted separately, sometimes even in different locations. Without interactive modeling, it is difficult to ensure that the acquired data are sufficient for a complete reconstruction, which in the case of error, significantly increases the cost and time required for modeling. Moreover, hand-held sensors are often designed using laser curves projected on the surface of an object. Nevertheless, methods for model reconstruction and pose refinement from range curves are nonexistent or too complex.
Approach: Our hypothesis consists in suggesting that the solution to 3D modeling problems does not solely depend on new methods, but rather on an adequate representation of shapes in 3D space. The representation that we have selected is an implicit representation in the form of a vector field. This enables the incremental integration of multiple views and contains all of the necessary information for an efficient registration whose complexity is linear with respect to the number of points measured on the surface. Vector fields can be efficiently compressed and displayed, which enables the performance of all of the modeling steps with a single representation of the surface.
Challenges: The main challenge of this project is the development of algorithms for all of the modeling steps: registration, reconstruction, compression and surface rendering using an implicit representation, i.e. vector fields. In particular, all of these algorithms must be incremental and of linear complexity to enable an interactive modeling. The algorithms must also enable the reconstruction of models from all types of range data: point clouds, curves measured on the surface and range images. They must also allow the combination of different types of data so as to build a model.
Applications: Even if this is a relatively new research area, 3D modeling has several applications such as virtual reality, reverse-engineering, industrial inspection, biometry, forensics and archaeology.
Calendar: September 2000 - December 2005
Last modification: Sep 28 2007 2:35PM by tdragan


©2002-. Computer Vision and Systems Laboratory. All rights reserved