Abstract: The theory of Particle Swarm Optimization (PSO) has been advanced considerably over the last two decades. A main focus was the convergence of the particles where different frameworks and assumptions have been adopted which lead to a variety of outcomes differing in their relevance as a predictor of the performance of the algorithm. We discuss here the interesting phenomenon that the study of the stochastic dynamics of the particles provides a limit case, while the average dynamics yields a different limit. The practically relevant parameter regions are well described by a combination of these two frameworks. In this way, we can also specify differences between the best particle and other particles in the swarm which has an effect on the performance of the algorithm and was not considered in previous approaches. In addition to benchmark results, also the swarm entropy confirms this formulation of the stability properties of PSO. The entropy effects become obvious when considering PSO as a stochastic differential equation. This framework enables us to describe the time scales of excursions of particles away from the best location found so far, which indicates the diversity of locations in the search space discovered so far by the particles.
Title: Unsupervised viewpoint estimation using category-level neural radiance fields
Abstract: Recovering camera parameters from images is usually done with structure-from-motion algorithms, but these are tied to a single scene and are unable to estimate viewpoint for a category of objects in a consistent way. Recent progress in 3D reconstruction has shown promising results in the form of neural radiance fields, allowing complex scenes to be modelled by a neural network. In this talk, we'll explore how to leverage these new tools in an analysis-by-synthesis approach to learn to predict category-level camera viewpoints in a fast and efficient manner.