Tracking for Learning an Object Representation from Unlabeled Data

For learning an object representation a huge amount of labeled data is needed. To minimize the labeling effort this paper proposes a new approach for learning from unlabeled data. The main idea is to combine a tracker and a learning method by directly feeding the learning algorithm with patches obtained by the tracker. In particular we apply an MSER based tracker and batch PCA for learning. But in general any tracker and any learning algorithm may be used. In a first step, the object-of-interest is initialized manually. Then, the object is tracked through a video sequence and a set of image patches, showing the object from different views, is extracted. The obtained patches are passed to PCA and a reconstructive model of the object is learned. Human input is reduced to a one-time initialization of the tracker. The approach is demonstrated on realistic scenes including face detection and detection of hand held objects.