Together with Stéphane Lathuilière, we will give a tutorial at ICIAP’19 on regression for computer vision. In this course we will first describe the basics of regression, probabilistic regression and deep regression. We will present recent results of these methodologies when tackling different computer vision applications such as pose or gaze estimation. We will then provide practical recommendations for using vanilla ConvNets for regression. After, we will provide details on recent methods combining deep architectures and probabilistic models for robust regression. Finally , we will present how generative models can be conditioned on continuous values. In particular, we will focus on adversarial approaches for high quality image generation. We will introduce the problem of generating person and face images conditioned on a given continuous value (e.g. pose or facial expression). In these cases, generation suffers from pixel-to-pixel misalignments and other perturbation problems that can be duly addressed with recent state-of-the-art methodologies. We will present how these approaches can be extended to video generation and, in particular, image animation.
Here are the slides of the tutorial: