PhD graduated
Team : MLIA
Departure date : 10/10/2019

Supervision : Matthieu CORD

Unsupervised Learning of 3D Shape Spaces for 3D Modeling

Even though the 3D media becomes increasingly more popular, especially with the democratization of virtual and augmented experiences, it remains very difficult to manipulate a 3D shape, even for designers or experts. Indeed, the creation or the editing of 3D contents require advanced skills in 3D modeling. Despite all the work done in computational geometry and 3D processing, the software dedicated to 3D still targets a professional audience due to the complexity of the 3D modeling task. Following the recent achievements of machine learning in the 2D domain, and especially the breakthrough of deep learning, several works have shown some success to adapt these learning methods to the 3D unsupervised learning setting. These techniques can be leveraged to provide intelligent 3D modeling and editing tools, with a user-friendly interface for inexperienced users as the eventual goal. This thesis falls within this research topic, and aims at learning convincing 3D representations, which could then be used to conceive new intelligent 3D modeling tools or within already exiting CAD software. Primarly, given a 3D database containing 3D instances of one or several categories of objects, we want to learn the manifold of plausible shapes, i.e. a latent representation generalizing the samples of the training dataset. However, the manifold of plausible shapes is often much more complex compared to the 2D domain. Indeed, 3D surfaces can be represented using various embeddings (voxels, distance fields, point clouds, meshes, structure trees, etc.), and may also exhibit different alignments and topologies. Besides, the manifold itself can be seen from different points of views, to uncover all the diversity of the 3D domain, each one leading to its own issues. In this thesis we study the manifold of plausible shapes in the light of the aforementioned challenges, by deepening three different premises. In chapter 3, we consider the manifold as a quotient space, in order to learn the shapes’ intrinsic geometry from a dataset where the 3D models are not co-aligned. In chapter 4, we assume that the manifold is disconnected, which leads to a new deep learning model able to automatically cluster and learn the shapes according to their typology. Finally, in chapter 5, we study the translation of an unstructured 3D input to an exact geometry, represented as a structured tree of continuous solid primitives. Where possible, we also include theoretical insights supported by formal theorems, to justify as much as possible the scientific rigor of our approaches.

Defence : 10/10/2019

2019 Publications