Deep learning for deformable objects
Deep convolutional neural networks (DCNNs) are currently the predominant paradigm for learning from labeled data in computer vision applications. For example DCNN architectures are the methods of choice for object recognition, detection, as well as semantic segmentation. In this talk, I will demonstrate how DCNNs can be used to model deformable objects. That is, I will start by briefly introducing the problem of statistical deformable model fitting using non-linear least squares and I will show how machine learning has reformulated the problem. Then I will demonstrate a general DCNN plus Recurrent Neural Network (RNN) architecture for solving non-linear least squares. I will take a step further and demonstrate how fully convolutional neural networks can be used for fitting deformable 3D models, as well as for estimating properties of 3D objects, e.g. surface normals.
Geometric deep learning for shape analysis
The past decade in computer vision research has witnessed the re-emergence of "deep learning" and in particular, convolutional neural network techniques, allowing to learn task-specific features from examples and achieving a breakthrough in performance in a wide range of applications. However, in the geometry processing and computer graphics communities, these methods are practically unknown. One of the reasons stems from the facts that 3D shapes are non-Euclidean spaces, hence the very notion of convolution is rather elusive. In this talk, I will show some recent works trying to bridge this gap. Specifically, I will show the construction of intrinsic convolutional neural networks on meshes and point clouds, with applications such as defining local descriptors, finding dense correspondence between deformable shapes and shape retrieval.