Unsupervised Hyperbolic Action Recognition

Abstract

Methods based on Deep Geometric Learning allow the development of solutions with a geometric approximation in different applications. In particular, the curved feature of hyperbolic space has the ability to describe hierarchical structures in a better manner. In this paper, we aim to define an unsupervised learning model for action recognition. The curved feature space is intended to be used to describe a hierarchical relationship between the clips that compose a complete video sequence. These, in turn, are related to each other by means of a triplet loss function and a VAE (Variational Auto-Encoder) neural architecture, which establishes a similarity relationship between clips to identify actions from a set of unlabelled data.