http://www.vlfeat.org/~vedaldi/

AG 1, AG 3, AG 5, SWS, AG 2, AG 4, RG1, MMCI   Info

MPI Audience

English

45 Minutes

Saarbrücken

Machine learning has had a central role in most of the recent advances in semantic image analysis, including image classification and object category detection. In fact, off-the-shelf classifiers, such as the ones used in our multiple kernels object detector, often achieve state-of-the-art performance in these tasks. Nevertheless, we argue that generic machine learning tools will not scale to the construction of general purpose vision systems due to the tremendous variability of the natural world and of its appearance. Simply increasing the capacity of the models is insufficient as the resulting computational cost is intolerable and learning is too inefficient to be possible from any reasonable amount of data.

The computational burden can sometimes be ameliorated by exploiting the mathematical structure of the model. As an example, we introduce the homogeneous kernel map, a powerful technique that, by "linearizing" common non-linear kernels, dramatically accelerates learning. Improving the statistical efficiency requires however modelling the deeper structure of the data. To this end we propose structured learning and demonstrate applications to object detection under partial truncations and efficient semantic segmentation based on holistic models of the segments. Finally, to reduce the cost of estimating complex structures, which is often combinatorial, we investigate using coarse-to-fine cascades, and apply them to real-time object detection.

Peter Gehler

--email hidden

passcode not visible

logged in users only