In recent years, the problem of content-based shape retrieval (CBIR) has attracted the interest of scientists. The objective of such system is to retrieve, from a given 3D query, the most similar 3D models from a given database; a linked issue is to find the correspondence (sparse or dense) between the query and the retrieved models. This problem is not easy since, to be really efficient, such retrieval/classification system has to be robust to common 3D shape variations like connectivity change, non-rigid deformation (isometry), local deformation or cropping.
In this thesis the objective is to propose robust and real-time partial shape retrieval algorithms, dedicated to graphics databases coming from the entertainment industry (gaming, motion pictures, etc.). The proposed system aims to be integrated into a whole assembly-based content creation software. To fulfill the requirements of this application, the proposed shape retrieval system will have to own the following properties:
Robustness to strong geometrical and topological changes (i.e. being able to recognize a model even if its pose has changed or even if its surface is not exactly the same as the query).
Handling of partial query (the most critical issue regarding our application).
Integration of metadata specific to entertainment industry graphics: Texture atlas, displacement maps, etc.
Reasonable computation time for the offline indexing part and real-time for the online retrieval step.
The proposed retrieval algorithms will be integrated with geometric transfer algorithms developed by our partners (LTCI, Télécom ParisTech) to produce a complete example-based 3D modeling system.