Embedding Gestalt Principles to Hierarchical Distance Dependent Nonparametric Bayesian Model for Video Segmentation

Abstract

Gestalt is a psychology term meaning “unified whole” which refers to the theories of visual perception developed in the 1920s. These theories attempt to describe how people tend to organize visual elements into groups or unified wholes when certain principles are applied. These principles such as similarity, common fate, continuation and etc. are intuitive to understand. However the challenge is to encode Gestalt theory as the principle to construct a computational model for visual process. In the domain of computer vision, visual processing tasks such as segmentation benefits greatly from spatio-temporal information from videos. Hence, we propose to study video segmentation where the number of objects is unknown. We achieve this by formulating a hierarchical nonparametric Bayesian model. Our model contains three key features 1) it embeds Gestalt principles as the prior of the model 2) it is a distance dependent nonparametric Bayesian model where the spatial temporal order of the data points matters. 3) it is a hierarchical model where we considered both the local and global aspects. We show that that our unsupervised generative model share similar results in human visual segmentation tasks as well as some psychology experiments.