Lab Meeting 6 August (Yu-Hsiang) : Unsupervised Activity Perception by Hierarchical Bayesian Models

Author: Xiaogang Wang, Xiaoxu Ma, Eric Grimson

Abstract :
We propose a novel unsupervised learning frameworkfor activity perception. To understand activities in complicatedscenes from visual data, we propose a hierarchicalBayesian model to connect three elements: low-level visualfeatures, simple “atomic” activities, and multi-agent interactions.Atomic activities are modeled as distributions overlow-level visual features, and interactions are modeled asdistributions over atomic activities. Our models improve existinglanguage models such as Latent Dirichlet Allocation(LDA) and Hierarchical Dirichlet Process (HDP) by modelinginteractions without supervision. Our data sets arechallenging video sequences from crowded traffic sceneswith many kinds of activities co-occurring. Our approachprovides a summary of typical atomic activities and interactionsin the scene. Unusual activities and interactions arefound, with natural probabilistic explanations. Our methodsupports flexible high-level queries on activities and interactionsusing atomic activities as components.

link