Properties of SIFT-descriptors

Some things to note

• These experiments were conducted on SIFT descriptors using Vedaldi's implementation, extracted from the Oxford buildings dataset.
• Note that the descriptors were extracted from Keypoints (as in Lowe's and Vedaldi's DOG implementations), and thus the statistics may differ from descriptors extracted randomly.

Dimensionality

• 300,000 descriptors were used (extracted from random images in the Oxford buildings dataset).