Local learning algorithms with application to action recognition and video analysis
Metadata[+] Show full item record
[ACCESS RESTRICTED TO THE UNIVERSITY OF MISSOURI AT AUTHOR'S REQUEST.] Activity analysis has been an active research area in recent years, due to its difficulties lying in the feature descriptor, video representation and learning algorithms. Several algorithms are proposed to conquer these difficulties. First, a mid-level feature, Histogram of Oriented Gradients Variations (HOGV), is developed for action recognition to solve the feature description problem. The proposed HOGV is not only stable due to its cell-block structure, but is also capable of capturing the static and dynamic characteristics of human actions. Second, video representation is one of the key problems in video analysis. To cope with the multi-label and rich context nature of videos, we propose to represent a video with a textual description resulted from video-to-text translation. This fully automatic translation is achieved by mapping local visual descriptors to key words distributions using a Visual-Textual Distribution (VTD) tree learned autonomously from the Internet. Third, we propose a new local learning algorithm noticing that rarely training data are evenly distributed in the input space, which downgrades the performance of the linear SVM classifier. Partitioning the input space in tandem with local learning may alleviate the unevenly data distribution problem. However, the extra model complexity introduced by partitioning frequently leads to overfitting. To solve this problem, we proposed Randomized Support Vector Forest (RSVF): Many partitions of the input space are constructed with partitioning regions amenable to the corresponding linear SVMs. The randomness of the partitions prevents the overfitting introduced by the over-complicated partitioning. Finally, we further explored the potential of non-patch feature. A generalized superpixelization algorithm with boundary preserving distance metric is proposed. It outperforms state-of-the-art superpixelization algorithms in three aspects: Generalizing to color and highly-textured images, generating compact superpixels, and computational efficiency.
Access is limited to the campus of the University of Missouri-Columbia.