An unsupervised approach to dominant video scene clustering
Haiping Lu, Yap‐Peng Tan · 2003
In this paper, we propose an unsupervised approach for dominant scene clustering in sports video. By adopting a customized peer group filtering (PGF) to identify prototypes for k-means clustering, dominant scenes can be clustered based on shot color histogram (SCH). Meanwhile, the number of clusters can automatically be determined by estimating the time coverage of dominant scenes. To improve the computational efficiency and clustering accuracy, principal component analysis (PCA) and linear discriminant analysis (LDA) are used to project SCH into reduced dimensional spaces. The prototypes obtained by PGF can also be served as sufficient and representative training data for LDA. Such good training data ensures LDA outperforms PCA with better clustering performance in more reduced feature dimension.