Modeling motion flow using tensor dynamic textures

As a family of visual patterns in moving scenes with certain temporal regularity, dynamic textures are powerful visual cues for people to understand things; hence, effective models are needed for relevant applications. Considering that image sequences are really tensor time series, this paper proposes a tensor dynamic texture model to represent dynamic texture videos, and a sub-optimal algorithm to estimate the model parameters. Our tensor-based method can capture multiple interactions and essential structures in videos. Experimental results on dynamic texture synthesis show that the proposed method not only achieved a better visual quality, but also a smaller model size and a less time cost. The maximum PSNR gain achieves 2.36 dB, and the maximum model size reduction achieves 49.68%.