Global image features for scene recognition invariant to symmetrical reflections in robotics

Scene understanding is still an important challenge in robotics. Robots must be aware of the kind of the environment where they move. In our case we plan to combine scene recognition with a multisensor localization system to allow the retrieval of knowledge (as robot controllers), according to where the robot is. In this paper we analyse the impact of several global image representations to solve the task of scene recognition. The performances of the different alternatives were compared using a benchmark of images taken in the Centro Singular de Investigacion en Tecnoloxias da Informacion (CITIUS), at the University of Santiago de Compostela, since this is the environment where the robot moves. The results are promising not only regarding the accuracy achieved, but mostly because we have found a holistic representation that allows the correct classification of images corresponding to rooms that are symmetrical (we increased the size of the test set including images that are obtained by specular reflection from other images also included in the same set).

keywords: Scene recognition, holistic representations, invariance to symmetries, CENTRIST, spatial pyramid, Local Difference Binary Patterns.