Performance Evaluation of State-of-the-Art Filtering Criteria Applied to SIFT Features

Feature matching is an important and crucial task in computer vision. Unfortunately, the features extracted from images are usually redundant or irrelevant, leading to many ambiguities and false positives during matching. This work provides an exhaustive study of state-of-the-art feature filtering strategies investigating different criteria, such as the contrast level, inner primary ratio, entropy, saliency and median value, in the purpose of discarding irrelevant features. The evaluation on the well-known Oxford-5k dataset demonstrates the relevance of these criteria to reduce the amount of data involved in the matching step and the number of false positives, thus leading to faster and more accurate matching with almost no loss in performance. Our results are also compared to convolutional neural networks (CNNs), providing extensive experiments that corroborate recently published work.