Predictive clustering for image annotation and retrieval

Saso Dzeroski (Jozef Stefan Institute, Slovenia)


Predictive clustering for image annotation and retrieval The predictive clustering paradigm unifies the predominant machine learning approaches of predictive modeling and clustering. It can solve a variety of problems of predicting structured outputs, such as multi-target regression ahd hierarchical multi-label classification. It elegantly handles different amounts of supervision in learning, ranging from fully supervised, via semi-supervised, to completely unsupervised learning.

We have used predictive clustering in bag-of-visual-words approaches to image annotation and retrieval. It was used to learn to annotate images in (hierarchical) multi-label settings. It was also used to construct visual codebooks, which were then used for image annotation and retrieval. This produced visual codebooks with superior discriminative power and thus better retrieval performance, while maintaining excellent computational efficiency.