Machine Learning and Probabilistic Models of Vision

This paper draws connections between Probabilistic Methods and Machine Learning Approaches. In particular, we show that many Support Vector Machine criteria – binary, multi-class, and latent – can be obtained as upper bound approximations to standard probabilistic formulations. The advantage of these ’Machine Learning bounds’ is that it greatly simplifies the computation and, possibly, may yield greater robustness. These connections enable us to take complex models formulated in terms of probabilistic distribution defined over graph structure and approximate/bound them by machine learning techniques. We illustrate this with examples from the literature which are applied to a range of vision problems — including image labeling, object detection and parsing, and motion estimation – and which achieve state of the art results.