A Framework for Classifier Adaptation for Large-Scale Multimedia Data

摘要: Machine learning techniques have been used extensively to build models for the analysis and retrieval of multimedia data. The explosion data on Web poses a great challenge such not simply because sheer volume, but also heterogeneity With from wide variety domains, trained one domain do generalize well other while at same time it is prohibitively expensive new each every due high cost labeling training examples. In this paper, we tackle in large-scale using cross-domain model adaptation better performance reduced human cost. Specifically, investigate problem adapting supervised classifiers or more source domains classifier target that has only limited labeled foundation our work general framework function-level based regularized loss minimization principle, which adapts by directly modifying its decision function. Under framework, can derive concrete algorithms plugging any regularization functions, among elaborate adaptive support vector machines (a-SVM). We further extend multiclassifier adaptation, namely multiple existing into domain, way contributions these are automatically determined. evaluate proposed approaches semantic concept detection TRECVID corpora. results show outperform (adaptation nonadaptation) methods terms accuracy and/or efficiency, offers benefits.