Computer Vision News - September 2016

34 Computer Vision News Research Research Qualitative Results: The authors conducted an impressive and exhaustive comparison to 16 other feature extraction methods, outperforming each one of them on each of the 3 datasets used: M. Score - Matching Score: the ratio of ground truth correspondences that can be recovered by the whole pipeline over the number of features proposed by the pipeline. This metric measures the overall performance of the pipeline. In the deep evaluation procedure, the metric used were as follows: to demonstrate on Strecha that each of the 3 LIFT layers is indeed essential, 2 LIFT models were evaluated (one trained on Piccadilly dataset and the other on Roman Forum dataset). In each model, the LIFT layer was interchanged with its SIFT counterpart, showing that each element of the pipeline is indeed crucial: The metrics used are: • Rep. (Repeatability of feature points): the ratio of key-points that are found consistently in the shared region. • NN mAP (Nearest Neighbor mean Average Precision): Area Under Curve (AUC) of the Precision-Recall curve, using the Nearest Neighbor matching. Results show that training the pipeline as a whole is important for optimal performance.