Traditionally, augmented reality systems place virtual objects in the real world using fiducial markers. Nowadays, in order to avoid such artificial markers that may become intrusive to the environment, markerless tracking techniques have been applied to the so named markerless augmented reality systems. The removal of the markers from the scene implies the use of more sophisticated computer vision algorithms, capable of describing what is being seen based uniquely on the image features. Many works have been performed focusing the optimization of such algorithms, but the choice of target platform and best optimization approaches to use is still difficult. This work intends to alleviate this choice by explaining the pros and cons of each target platform, and highlighting the experience obtained while developing augmented reality algorithms for each one of them.