Domain Transfer SVM for Video Concept Detection

1. Lixin Duan Ivor W. Tsang Dong Xu Nanyang Technological University Stephen J. Maybank University of London • Kernel function based on multiple base kernels: • SVM decision function: • is any monotonic increasing function and is a tradeoff parameter. where • is a column vector with entries, in which the first entries are set as and the • remaining entries are set as respectively. . Dual Form • Label vector: • Kernel matrix of the labeled samples: • Auxiliary domain with samples. • Target domain with samples. • Compare data distributions based on the square distance between the means of samples from the two domains in the Reproducing Kernel Hilbert Space (RKHS). . • Unlabeled data can be employed. . Domain Transfer SVM for Video Concept Detection Experiments Introduction Domain Transfer SVM • Data Sets • Predicting Method • In this paper, we propose a novel semi-supervised cross-domain kernel learning method, referred to as Domain Transfer SVM (DTSVM),for the challenging video concept detection task to deal with the tremendous change of keyframe features distributions from different domains. • DTSVM simultaneously learns a kernel function and a robust SVM classifier by minimizing both the structural risk functional of SVM and the distribution mismatch of labeled and unlabeled samples between the auxiliary and target domain. . • we assume that the kernel function in SVM learning is from a linear combination of multiple base kernels. . • we propose an efficient learning algorithm to solve the linear combination coefficients of kernels and the SVM classifier under a unified convex optimization framework. . • Group 1 and 2: DTSVM_AT • Group 3: DTSVM_T • TRECVID 2005: 61,901 keyframes from 108 hours of video programs • TRECVID 2007: 21,532 keyframes from 60 hours of news magzines, etc. • Three Low-Level Global Features • Grid Color Moment (225 dim.) • Gabor Texture (48 dim.) • Edge Direction Histogram (73 dim.) • First Objective To form a 346-dimensional feature vector • Four Types of Kernels • Gaussian: • Second Objective • Laplacian: Maximum Mean Discrepancy • Inverse Square Distance: • Inverse Distance: • Final Formulation • Employ multiple base kernels • Quadratic in and , respectively. • Solved by updating and iteratively. Computer Vision and Patter Recognition 2009