Singular Patterns in Optical Flows as Dynamic Texture Descriptors

Abstract

This work introduces a novel approach to dynamic texture description. The proposed method is based on statistics from a vector field feature extractor that decomposes and describes features of distinctive local vector patterns as composites of singular patterns from a dictionary. The extractor is applied to a time-varying vector field, namely a dynamic texture’s optical flow frames. An interest point pooling method statistically highlights the recurring texture patterns, generating a histogram signature that is descriptive of the temporal changes in the texture. The proposed descriptor is used as feature vector on classification experiments in a widespread dataset. The classification results demonstrate our method improves on the state of the art for dynamic textures with non-trivial motion, while employing a smaller feature vector.