Thesis of Vincent Leon

Semantic 3D model description – Application to example-based modeling

Even though there are more and more applications using 3D models, creating 3D models themselves is a time-consuming process. Example-based modeling makes this process more efficient by enabling the user to use existing model and assemble them, in order to obtain a new model. To use these models, it is necessary to find them easily in a collection using partial mesh retrieval techniques. The geometry of each part must then be adapted, before assembly. State-of-the-art methods focus on segments assembly mechanisms: database models are previously segmented and labeled. This hypothesis prevents several manual segmentations, but it also limits the creative process as a series of additions and replacements of pre-determined parts. In this thesis, modeling is done as replacement of parts from a original model, but these parts are not already segmented. We present a new continuous semantic descriptor computed on the whole 3D mesh, to find among the database models the parts corresponding to a user’s selection on the reference model : if the user selects the elbow of a humanoid character, all the elbows of the database will be returned, even though the label is not explicitly used for labeling the database. Thanks to this continuous semantic description and this patch correspondance mechanism, we can get all the necessary parts to create a new model. These different parts are then assembled using a soft registration technique and remeshing.

Jury

Directeurs de thèse : Jean-Philippe Vandeborre, Guillaume Lavoué Rapporteurs : Géraldine Morin, Laurent Lucas Examinateur : François Goulette

Thesis of the team 3D SAM defended on 10/11/2016