ABSTRACT
Contents 2.1 Survey of Cross-Domain Learning for Concept Detection . . . . . . . . . . . . . . . . . 29
2.1.1 Standard SVM.. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30 2.1.2 Semi-Supervised Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 2.1.3 Feature Replication . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 2.1.4 Domain Adaptive Semantic Diffusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 2.1.5 Adaptive SVM (A-SVM) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33 2.1.6 Overview of Our Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
2.2 CDSVM for Data Incorporation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33 2.3 AS3VM for Incremental Classifier Adaptation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
2.3.1 Discriminative Cost Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 2.3.2 Graph Regularization on Data Points . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 2.3.3 Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
2.4 Prediction-Based Concept Score Incorporation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39 2.4.1 Overview of Our System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40 2.4.2 ELF Learning Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42 2.4.3 Semantic Concept Classification with Multitype ELFs . . . . . . . . . . . . . 43
2.4.3.1 Cross-Domain ELFs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43 2.4.3.2 Within-Domain ELFs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44 2.4.3.3 Classification with ELFs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
2.5 Experiments: Cross-Domain Learning in TV Programs . . . . . . . . . . . . . . . . . . . . 45 2.6 Experiments: From TV Programs to Consumer Videos . . . . . . . . . . . . . . . . . . . . 47
2.6.1 AS3VM for Semantic Concept Detection. . . . . . . . . . . . . . . . . . . . . . . . . . . 47 2.6.2 Prediction-Based Method for Concept Detection in Events. . . . . . . . 49
2.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50 Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52 References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
Automatic semantic concept detection has become increasingly important to effectively index and search the exploding amount of multimedia content, such as those from the web and TV broadcasts. The large and growing amount of unlabeled data in comparison with the small amount of labeled training data limits the applicability of classifiers based upon supervised learning. In addition, newly acquired data often have different distribution from the previous labeled data due to the changing characteristics of real-world events and user behaviors. For example, in concept detection tasks such as TRECVID [19], new collections may be added annually from unseen sources such as foreign news channels or audiovisual archives. There exists a non-negligible domain difference. To improve the semantic concept detection performance, these issues need to be addressed.
