Scene recognition is an important topic in mobile robotics; nevertheless, it is still difficult to reach robust and reliable solutions, in fact there ar many studies which have clearly arrived at the conclusion that human visual scene recognition is view point dependant. This paper suggests a metric aimed to value the images according to their coverage and thus anlayse whether they contain enough information to identify the scene. This metric prioritizes images which contain key points scattered all over, and which reflect the presence of close and far obstacles. The experimental results show to what extent the metric described in this paper resembles what the human would choose as relevant. In particular this metric was used to rank the images from two datasets and the results were compared with the selection made by humans who chose the most representative ones.
Keywords: scene recognition, robotics, relevance and view-point.