Online appearance manifold learning for video classification and clustering

Abstract

Video classification and clustering are key techniques in multimedia applications such as video segmentation and recognition. This paper investigates the application of incremental manifold learning algorithms to directly learn nonlinear relationships among video frames. Video frame classification and clustering are performed to the projected data in an intrinsic latent space. This approach has avoided partitioning video frames into arbitrary groups. It works even when the input video frames are under-sampled or unevenly distributed. Experiments show that video classification and clustering give better results in the latent space than in the original high dimensional space.