We present a real-time localization system based on monocular vision and natural landmarks. In a learning step, we record a reference video sequence and we use a structure from motion algorithm to build a model of the environment. Then in the localization step, we use this model to establish correspondences between the 3D model and 2D points detected in the current image. These correspondences allow us to compute the current camera localization in real-time. The main topic of this paper is the performance evaluation of the whole system. Four aspects of performance are considered : versatility, accuracy, robustness and speed.