Support vector machine active learning for image retrieval
MULTIMEDIA '01 Proceedings of the ninth ACM international conference on Multimedia
Convex Optimization
Evaluation campaigns and TRECVid
MIR '06 Proceedings of the 8th ACM international workshop on Multimedia information retrieval
Manifold Regularization: A Geometric Framework for Learning from Labeled and Unlabeled Examples
The Journal of Machine Learning Research
Kodak's consumer video benchmark data set: concept definition and annotation
Proceedings of the international workshop on Workshop on multimedia information retrieval
Large-scale multimodal semantic concept detection for consumer video
Proceedings of the international workshop on Workshop on multimedia information retrieval
Correlative multi-label video annotation
Proceedings of the 15th international conference on Multimedia
Cross-domain video concept detection using adaptive svms
Proceedings of the 15th international conference on Multimedia
Semi-Supervised Learning
Proceedings of the 1st ACM International Conference on Multimedia Retrieval
WSM2011: third ACM workshop on social media
MM '11 Proceedings of the 19th ACM international conference on Multimedia
Multi-modal region selection approach for training object detectors
Proceedings of the 2nd ACM International Conference on Multimedia Retrieval
Hi-index | 0.00 |
Practical semantic concept detection problems usually have the following challenging conditions: the amount of unlabeled test data keeps growing and newly acquired data are incrementally added to the collection; the domain difference between newly acquired data and the original labeled training data is not negligible; and only very limited, or even no, partial annotations are available over newly acquired data. To accommodate these issues, we propose a Laplacian Adaptive Context-based SVM (LAC-SVM) algorithm that jointly uses four techniques to enhance classification: cross-domain learning that adapts previous classifiers learned from a source domain to classify new data in the target domain; semi-supervised learning that leverages information from unlabeled data to help training; multi-concept learning that uses concept relations to enhance individual concept detection; and active learning that improves the efficiency of manual annotation by actively querying users. Specifically, LAC-SVM adaptively applies concept classifiers and concept affinity relations computed from a source domain to classify data in the target domain, and at the same time, incrementally updates the classifiers and concept relations according to the target data. LAC-SVM can be conducted without newly labeled target data or with partially labeled target data, and in the second scenario the two-dimension active learning mechanism of selecting data-concept pairs is adopted. Experiments over three large-scale video sets show that LAC-SVM can achieve better detection accuracy with less computation compared with several state-of-the-art methods.