Thesis of Halim Benhabiles
Defense date: 10/08/2001
Advisor: Guillaume Lavoué
The three-dimensional models are generally represented as meshes of polygons (generally triangles). This kind of representation has the advantage of being perfectly adapted to 3D display with the help of modern 3D accelerated hardware. But the main drawback of this format is the lack of a structure or a hierarchical description that could be very useful for the applications cited above. Hence, the automatic segmentation of 3D-mesh models is very often a necessary pre-processing tool for these applications. Mesh segmentation consists in subdividing a polygonal surface into patches of uniform properties either from a strictly geometrical point of view or from a perceptual / semantic point of view.
To bring a solution to this problem, many systems were and are still currently developed for the segmentation of bidimensional data (images or videos). However these solutions are not really effective or not easily adaptable to intrinsically three-dimensional data. Moreover, one could easily notice that, contrary to the 2D-data domain, there is neither protocol, nor standard data collection for the comparison and the evaluation of the 3D segmentation methods.
In this context, and within the framework of the MADRAS project the goals of my thesis are the following three:
Building a collection of 3D mesh models – static 3D-models – with a ground-truth. The ground-truth will be composed of one or more segmentations for each of the 3D-model. These segmentations will come from hand-made segmentations and also from automatic or semi-automatic methods.
Exploiting the human factor in order to improve the conception and the evaluation of segmentation algorithms, through subjective experiments. The subjective and perceptual aspects will be used to build a reference toolkit that will allow an entire automatic comparison process of existing and future segmentation methods
Designing new segmentation algorithms for static 3D-mesh models, using the human factor, with machine learning techniques.
With this triple goal, the MADRAS project aims at helping the scientific communities involved in 3D-model segmentation. Such a benchmarking tool will allow the researchers to evaluate and compare existing and new segmentation methods. Moreover, the introduction of the human factor in segmentation methods, with subjective and perceptual aspects, is the first attempt in the 3D domain.