Robust Building Identification for Mobile Augmented Reality

Mobile augmented reality applications have received considerable interest in recent years, as camera equipped mobile phones become ubiquitous. We have developed a “Point and Find” application on a cell phone, where a user can point his cell phone at a building on the Stanford campus, and get relevant information of the building on his phone. The problem of recognizing buildings under different lighting conditions, in the presence of occlusion and clutter, still remains a challenging problem. Nister’s Scalable Vocabulary Tree (SVT) [1] approach has received considerable interest for large scale object recognition. The scheme uses heirarchical k-means to create a vocabulary of features or “visual words”. We first show how we can use a SVT and an entropy-based ranking metric to achieve 100% recognition on the well known ZuBuD data set [2]. We present a SVM kernel-based extension to the SVT approach and show that it achieves a 100% recognition rate as well. We discuss the shortcomings of the ZuBuD data set, and present a more challenging Stanford-Nokia data set, with promising results.