Slightly Supervised Learning of Part-Based Appearance Models

We extend the GMM-based approach of [Selection of scale-invariant parts for object class recognition], for learning part-based appearance models of object categories, to the unsupervised case where positive examples are corrupted with clutter. To this end, we derive an original version of EM which is able to fit one GMM per class based on partially labeled data. We also allow ourselves a small fraction of un-corrupted positive examples, thus obtaining an effective, yet cheap, slightly supervised learning. Proposed technique allows as well a saliency-based ranking and selection of learnt mixture components. Experiments show that both the semi-supervised GMM fitting with side information and the component selection are effective in identifying salient patches in the appearance of a class of objects. They are thus promising tools to learn class-specific models and detectors similar to those by Weber et al.[Unsupervised learning of models for recognition], but at a lower computational cost, while accommodating larger numbers of atomic parts.