Learning texton models for real-time scene context

We present a new model for scene context based on the distribution of textons within images. Our approach provides continuous, consistent scene gist throughout a video sequence and is suitable for applications in which the camera regularly views uninformative parts of the scene. We show that our model outperforms the state-of-the-art for place recognition. We further show how to deduce the camera orientation from our scene gist and finally show how our system can be applied to active object search.