Visual event recognition in videos by learning from web data

treeKL: A distance between high dimension empirical distributions

“…Using the distance as a kernel, one can also use SVM, which are usually more 195 efficient in such case [9,11,22,10].…”

“…Depending on the application, d(x 1 , x 2 ) can refer to the Euclidean distance [9] or to the χ 2 distance [22,10], etc. We use the distance proposed in equation (1)…”

“…In many application fields, in particular in computer vision, very efficient techniques rely on the idea of bag-of-features (bof), which model the empirical distributions with distribution over clusters computed adaptively from data [9,11,22,10].…”

Watching Unlabeled Video Helps Learn New Human Actions from Very Few Labeled Snapshots

“…Domain adaptation can improve event learning across two domains of videos, such as Web videos to consumer videos [10] or one benchmark dataset to another [5], and it can also help train an object detector from videos [21]. A novel technique to use multiple source domains and a mix of static and dynamic features is developed in [9].…”

Enhancing Action Recognition by Cross-Domain Dictionary Learning

“…The former one utilized the Adaptive Support Vector Machines (A-SVMs) to adapt one or more existing classifiers of any type to a new dataset, and the latter proposed a Domain Transfer Multiple Kernel Learning (DTMKL) method to simultaneously learn a kernel function and a robust SVM classifier by minimizing both the structural risk function of SVM and the distribution mismatch of labeled and unlabeled data in different domains. Duan et al [6] considered to leverage large amounts of loosely labeled web videos for visual event recognition using the Adaptive Multiple Kernel Learning (A-MKL) to fuse the information from multiple pyramid levels of features and cope with the considerable variation in feature distributions between videos across two domains.…”