Lecture 16: Learning Object Models for Detection

This lecture describes how to learn models for detecting objects. These models explicitly represent objects parts and allow the spatial relations between these parts to vary. This differs from the methods we used to model faces and text which did not represent parts and instead relied on features that were invariant to these spatial variations. But many objects have large spatial variations and it is not possible to find invariant features. Moreover, modeling the subparts explicitly enables us to detect them and hence parse the image to detect the positions of the parts (i.e. say where different parts of the object are – and not simply detect the object).