Duration: 58:22

867 views

11

I will present recent work from SIGGRAPH and CVPR. The first builds a generative model of fonts in an automatic and unsupervised learning process; the only input is a collection of existing font files and the output is a probabilistic manifold that can be used to create new typefaces. The second piece of work generalises probabilistic PCA and Active Appearance Models to overcome a fundamental weakness; existing subspace models are unable to model image datasets that cannot be readily aligned. Instead, we learn a subspace model in a new context space, a deterministic function of an input “part map”, that implicitly encodes correspondence and thus brings the data into alignment. We illustrate the value of this approach by considering two example tasks: structured in-painting and appearance transfer.