Much interest has been devoted to augmented reality (AR) in both academic and enterprise environments, and rightly so. AR enhances interaction with real-world objects by adding computer-generated information to these objects relevant to the very interaction taking place in that moment. Applications already exist, but all share a common flaw: in outdoor use they become slow and resource hungry due to the explosive growth in the number of parameters they have to process.

This paper presents an AR application, called environment dedicated descriptor (EDD), which promises to behave better in such environments. Like its existing counterparts, it is still based on a registration module coupled with a tracking module, but it turns out to be more resource friendly because it focuses first on the most computable and information-rich elements on a scene and then extracts semantic features from them with a random forest classifier; it can be trained to do so with machine learning algorithms. Because it is this classifier that improves the overall application performance, the whole paper is built around it.

The paper starts by describing the classifiers most widely used today. Then it presents this new classifier, called random forest classifier because it classifies object features in a series of binary trees, each tree dedicated to specific features, and then combines trees into forests. The paper describes the classifier both in formal logic terms and with pseudocode, and then tests it against the other classifiers mentioned in previous sections for computation speed, accuracy, and invariance to illumination, among other parameters.

The paper shows, with extensive use of graphs and tables, that EDD almost always performs better than the competitors (of course!). To be fair, it also recognizes EDD’s main weakness: in order to reach such excellent performance, EDD must be trained for the specific context in which it is to be used; otherwise, its performance is not so good.