Structured Regression for Efficient Object Detection

Christoph Lampert
Max Planck Institute for Biological Cybernetics, Tuebingen, Germany

A crucial step on the road towards automatic image interpretation is the identification and localization of objects in a scene, i.e. Object Detection. Most current methods for this task are modifications of Object Classification techniques, based e.g. on a sliding window procedure. This has two drawbacks: it is unclear how to best train a classifier for this task, and the large number of possible subwindows makes the problem computationally expensive.

I will explain how recent developments in machine learning allow us to overcome both problems by treating object localization as a regression task into a structured output domain. Using a maximum-margin formulation, we derive an algorithm that is trained and evaluated as a generalized support vector machine. This allows robust and efficient object detection for arbitrary object categories.