Action Recognition Karthik Prabhakar UCF REU 2008, Week 2 Report May 30, 2008
Research Topic • Action Recognition in videos • (ala Mikel’s presentation on Wednesday) • Goals • Perform under unconstrained environments • Achieve (semi) time-invariance
Summary of Niebleset al. • “Unsupervised Learning of Human Action Categories using Spatio-Temporal Words” • Video sequences represented as collection of spatio-temporal ‘words’. • Using pLSA and LDA, the system learns the probability distributions of the spatial-temporal words and the topics corresponding to human action categories.
LDA (Latent Dirichlet Allocation) • A generative-model that allows a set of observations to be explained by unobserved groups (these explain why some parts of the data are similar). Priors Video sequences Action categories Spatio-temporal words
Goals/Tasks for next week • Finalize a set of descriptors/features • Implement/Get DTM • Implement the system • (hopefully) Get some test videos running!