Feature Ranking in Dynamic Texture Clustering

The dynamic texture clustering over space and time, which considers a video as a sample from a linear dynamical system, has recently received great attention. While there exist many algorithms for time-varying characteristics and phenomena clustering, an important and challenging issue arising from these studies concerns is that how to compare the importance of each feature, and which features of the data should be used for time-varying clustering. In view of the aforementioned observations, a new unsupervised feature ranking for dynamic texture clustering is proposed in this paper. Different from previous works, where the selection criteria are often used to identify relevant features, our method has an ability to score relevant features, and allow us to evaluate the importance of each feature. Also, it is intuitively appealing, avoiding any combinatorial search. Numerical experiments on various time-varying characteristics and phenomena demonstrate the accuracy and effectiveness of the proposed method.