Learning based approaches with applications to vision tasks
Zhuowen Tu, Jiayan Jiang · 2011
In this work, we tackle some low-level and high-level vision problems with effective and yet efficient learning approaches. In particular, for image registration we propose a coarse-to-fine learning-based approach which dramatically reduces the search space of possible geometrical transformations. For image segmentation, a data assisted output coding scheme is proposed to handle the problem of classifying a huge number of categories a pixel might belong to. For part-based object detection, we consider a weakly supervised learning approach to circumvent the limitation and ambiguity intrinsic to manual annotations. For image retrieval, a self-smoothing operator is introduced to improve upon an initial similarity measure which better respects the data manifold structure. Our learning based approaches have been shown to either outperform existing methods in accuracy, or obtain comparable results to the state-of-the-art at a much lower computational overhead, or address a practical challenge with data annotation.