Feature-based Pose Estimation in Augmented Reality Applications



We explore the performance of a number of feature-based pose estimation schemes, which are combined and concatenated with feature detectors, feature descriptors, outlier removal algorithms, and PnP algorithms. To this end we design a series of methods to evaluate the performance in each stage. We find that the SIFT and SURF detectors can localize the feature more accurately than the FAST-like detectors. Though the Affine-SIFT is quite computationally expensive, it is extremely robust to large tilts. RANSAC+P3P is more reliable than other outlier-rejection schemes, nevertheless it needs more computation time. In the pose estimation stage, OPnP is perhaps the best one among state-of-the-art PnP algorithms. To our best knowledge, if a device can be applied with powerful compute capability like cloud computing, the optimal solution to address the feature-based pose estimation problem at the current stage is the combination of ASIFT, RANSAC+P3P, and OPnP.

  1. Introduction to AR
  1. Introduction to Detectors and Descriptors
  1. BRIEF and ORB and BRISK
  1. OI and RPP and EPnP and OPnP