A 3D feature model for image matching

The proposed algorithm identifies whether or not a test photo belongs to a set of co-located training images based on its spatial proximity to the training set. We leverage concepts from Lowe´s SIFT and Snavely´s Photo Tourism algorithms, and match an image by its 2D features to the 3D features representing the training set. To reduce complexity and increase efficiency, the proposed algorithm implements a compact representation of the image set by merging collections similar features. Test images are then matched with the derived structure. Finally, a decision statistic is determined based on the percentage of features that match. Receiver operating characteristics, computational analysis, and distributions are included in the performance analysis.