Exploiting High Dimensional Video Features Using Layered Gaussian Mixture Models

Analysis of video data usually requires training classifiers in high dimensional feature spaces. This paper proposes a layered Gaussian mixture model (LGMM) to exploit high dimensional features for classifying various shots in video. LGMM decomposes a high dimensional feature space by building a pyramid structure and estimating the distribution of local partitions in each layer using Gaussian mixtures from the bottom of the pyramid to the top. We reduce the dimension of features in each local region at a lower layer by projecting them onto the estimated Gaussian components. These projected feature vectors are then used to estimate the Gaussian mixture models at a upper layer. The final dimension of the feature is adjustable by choosing the number of Gaussians at the top layer of the pyramid. We demonstrate the proposed method using motion features to classify video shots. The proposed method is independent from low level features and can be extended to other classification tasks