Conditional Random Fields for Object Localization
Master's defence by Andreas Danielsen and Andreas Eilschou
It takes place September 13, at 13:45 in Meeting Room A + B, 2-0-04/06
We propose a probabilistic model of the object localization problem by formulating it in the framework of conditional random fields (CRFs) which allows parameter learning by maximum conditional likelihood (MCL) estimation. The true log-likelihood and its gradient are infeasible to compute even for medium-sized images and datasets and we consider four approximate methods for learning the parameters of the CRF model: stochastic gradient descent, contrastive divergence, pseudolikelihood and piecewise training.
Our contribution is threefold: First, we formulate and derive the true log-likelihood objective and its approximations for the object localization problem. Secondly, we show that the approximate learning methods obtain results comparable with exact MCL learning but only contrastive divergence, pseudolikelihood and piecewise training are feasible in practice. Thirdly, we show that our results are comparable with results obtained by a structured support vector machine.
Supervisors: Christoph H. Lampert (IST Austria), Christian Igel (DIKU)
Censor: Rasmus Larsen (DTU Informatik)