Gaussian Processes for Shape Modelling: a Probabilistic Registration Approach
Shape models find increasing applications in fields ranging from medicine to security or animation. Gaussian Processes (GP) have emerged as a powerful framework for developing 3D shape models and performing related tasks in a unified manner. However, existing methods rely on the availability of a large and complete dataset, which can be costly and time-consuming to acquire, often requiring extensive manual input and application-specific pre-processing methods. A more common setting is a dataset with large regions of missing data and outliers, where standard techniques have limited performance. I will discuss a new registration/fitting method derived within the GP framework, which is designed to overcome these challenges. By drawing a parallel with state-of-the-art probabilistic registration algorithms, the proposed method provides a more principled approach and achieves better results in a realistic dataset.
