This paper presents a new framework to evaluate feature descriptors on 3D datasets. The proposed method employs the approximated overlap error in order to conform with the reference planar evaluation case of the Oxford dataset based on the overlap error. The method takes into account not only the keypoint centre but also the feature shape and it does not require complex data setups, depth maps or an accurate camera calibration. Only a ground-truth fundamental matrix should be computed, so that the dataset can be freely extended by adding further images. The proposed approach is robust to false positives occurring in the evaluation process, which do not introduce any relevant changes in the results, so that the framework can be used unsupervised. Furthermore, the method has no loss in recall, which can be unsuitable for testing descriptors. The proposed evaluation compares on the SIFT and GLOH descriptors, used as references, and the recent state-of-the-art LIOP and MROGH descriptors, so that further insight on their behaviour in 3D scenes is provided as contribution too.
