Appearance modeling using passive vision
Philippe Lambert
Patrick Hébert (Supervisor)
Problem: The objective is to study the problem of modeling the appearance of an object using passive vision. Appearance modeling includes all processes which aim to obtain a digital model allowing the image of an object to be reproduced from any viewpoint (we are considering static objects under a fixed light source). Passive vision differs from active vision in that it does not involve contact with the modeled object and does not use controlled light sources. Only a series of calibrated photographs are used.
Approach: The two methods currently used in modeling the appearance of objects through passive vision are briefly compared here. Geometric approaches determine the shape of an object based on assumed reflectance properties of the object's surface. Image-based approaches use sampling and the reconstruction of the plenoptic function. There is, however, a certain confusion in the literature concerning which method should be categorized as an image approach. Our position to avoid this confusion is simple. An image-based approach is not based on assumed reflectance properties. For example, as soon as a correspondence between images is established, this cannot be referred to as an image-based approach, since establishing a match between images is only possible when the approach is based on a particular reflectance model (usually Lambertian).

Let us briefly recall the advantages and disadvantages of the two approaches. The advantage of image-based approaches is that they are not based on particular reflectance models. Geometric approaches, however, directly allow a change in the lighting conditions under which the object is modeled. Geometric approaches also require less data, thus requiring less photographs of the object and less memory space for the representation of the model.

This research project aims to clearly formulate the problem of modeling the appearance of an object through passive vision. Both geometric and image-based approaches will be investigated. It will be shown that geometric approaches are disadvantaged since their objective of obtaining the true shape of an object is not realistic when using passive vision. An image-based approach will thus be favoured, however, this implies the loss of the advantages of a geometric approach. Thus, it will be shown that the geometric approach does not offer advantages, but instead offers strong hypotheses.

Calendar: May 2005 -
Last modification: 2007/11/14 by plambert


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