Normalized Cuts and Image Segmentation
IEEE Transactions on Pattern Analysis and Machine Intelligence
Content-Based Image Retrieval Using Multiple-Instance Learning
ICML '02 Proceedings of the Nineteenth International Conference on Machine Learning
Multiple-Instance Learning for Natural Scene Classification
ICML '98 Proceedings of the Fifteenth International Conference on Machine Learning
MILES: Multiple-Instance Learning via Embedded Instance Selection
IEEE Transactions on Pattern Analysis and Machine Intelligence
Correlative multi-label video annotation
Proceedings of the 15th international conference on Multimedia
Weakly Supervised Object Localization with Stable Segmentations
ECCV '08 Proceedings of the 10th European Conference on Computer Vision: Part I
Correlative linear neighborhood propagation for video annotation
IEEE Transactions on Systems, Man, and Cybernetics, Part B: Cybernetics
CVPR'04 Proceedings of the 2004 IEEE computer society conference on Computer vision and pattern recognition
| Hi-index | 0.00 |
In this paper, we have addressed two key issues for leveraging large-scale loosely-tagged images for object classifier training: (a) loose object tags, e.g., multiple object tags are loosely given at the image level without identifying object locations in the images; (b) missing object tags, e.g., some object tags are missed and thus negative bags may contain positive instances. To address both the issues of loose object tags and missing object tags jointly, a novel multiple instance learning (MIL) algorithm is developed and it consists of the following key components: (1) An agreement value is defined for characterizing instance-tag relatedness; (2) Automatic instance clustering is performed and inter-cluster correlations are leveraged for agreement value refinement; (3) An automatic instance-tag alignment algorithm is developed for assigning multiple object tags into the most relevant image instances with maximum agreement values and determining uncertain image instances whose object tags are not available on the tag list (missing object tags); (4) Object co-occurrence contexts are leveraged to predict missing object tags for the uncertain image instances. Our experiments on large-scale loosely-tagged images have provided very positive results.