Log-Linear Mixtures for Object Recognition

We present the log-linear mixture model as a fully discriminative approach to object category recognition which can, analogously to kernelised models, represent non-linear decision boundaries. This model is applied to the problem of recognising object classes in natural images, which is one of the most fundamental and best researched problems in computer vision. Similarly to many recent approaches our method uses local image descriptors and learns an object model from weakly annotated training data (i.e. only class labels). The use of local image descriptors has become a de-facto standard, because it has several advantages: