The use of three-dimensional models in the multimedia applications, is growing both in number and in size. The development of mode- ling tools, 3D-scanners, graphic accelerated hardware, Web3D and so on, offers access to three-dimensional materials of high quality. The constantly increasing needs concerning these kinds of data, are rapidly changing. While it becomes more and more easy to create new 3D-models, what about process and analysis after the creation of the 3D-models? Today, the 3D designer no longer asks : "How to create a new 3D-model?", but more probably "How to retrieve an existing 3D-model that is similar to those I already own in order to reuse it?" and "How to get the inner structure of a 3D-mesh model without any a priori knowledge on it?" This habilitation thesis aims to provide some answers to these two questions. In response to the first question, we developed a new Bayesian framework to retrieve 3D-models from a query made of one or more 2D- views, or of an entire 3D-model. The framework has been tested in an industrial application context and with an international benchmark. Each of these experiments has shown excellent results. The second question has been addressed in terms of topological analysis of the 3D-meshes with the help of Reeb graphs. This theoretical work has been applied to several practical domains, such as automatic 3D-mesh deformation, 3D-model retrieval, and 3D-mesh segmentation, and has always highlighted outstanding results. Finally, the segmentation of 3D-meshes, which is a frequent pre-processing step before any other analysis of the mesh, has drawn our attention. We proposed a reliable and robust metric to compare segmentations and evaluate the performances of the 3D-mesh segmentation methods, as well as a new learning-based segmentation approach that out- performs existing ones. To conclude, new perspectives of research on 3D- meshes are open.
