Feature Extraction

Revision as of 05:30, 16 October 2008 by Axtimwalde (talk | contribs) (Scale Invariant Interest Point Detector)
MOPS feature correspondences (example 1)
MOPS feature correspondences (example 2)

The plugins "Extract SIFT Correspondences" and "Extract MOPS Correspondences" identify a set of corresponding points of interest in two images and export them as PointRoi. Interest points are detected using the Difference of Gaussian detector thus providing similarity-invariance. Corresponding points are best matches from local feature descriptor that are consistent with respect to a common geometric transformation.

The plugins use the Scale Invariant Feature Transform (SIFT) and Multi-Scale Oriented Patches (MOPS) for local feature description. The thus established matches are filtered using the Random Sample Consensus (RANSAC).


SIFT parameters

Scale Invariant Interest Point Detector

initial gaussian blur
Accurate localization of keypoints requires initial smoothing of the image. If your images are blurred already, you might lower this value slightly to get more but eventually less stable keypoints. Increasing this value increases the computational cost for Gaussian blur, setting it to 3.2 is equivalent to keep 1.6 and use half maximum image size. Tip: Keep the default value 1.6 as suggested by Lowe (2004).
steps per scale octave
Keypoint candidates are extracted at all scales between maximum image size and minimum image size. This Scale Space is represented in octaves each covering a fixed number of discrete scale steps from σ to 2σ. More steps result in more but eventually less stable keypoint candidates. Tip: Keep 3 as suggested by Lowe (2004) and do not use more than 10.
minimum image size
The Scale Space stops if the size of the octave would be smaller than minimum image size. Tip: Increase the minimum size to discard large features (i.e. those extracted from looking at an image from far, such as the overall shape).
maximum image size
The Scale Space starts with the first octave equal or smaller than the maximum image size. Tip: By reducing the size, fine scaled features will be discarded. Increasing the size beyond that of the actual images has no effect.
upscale image first
Create the first scale octave with double the size of the original image in order to find features at the original pixel resolution. Tip: Do this only for very small images and if you desperately need more features.

SIFT correspondences

MOPS correspondences


  1. Lowe, David G. (2004). "Distinctive Image Features from Scale-Invariant Keypoints". International Journal of Computer Vision 60 (2): 91-110. doi:10.1023/B:VISI.0000029664.99615.94.
  2. Brown, Matthew; Szeliski, Richard; Winder, Simon (2005). "Multi-Image Matching Using Multi-Scale Oriented Patches". Proceedings of the 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05) Volume 1: 510-517. doi:10.1109/CVPR.2005.235.
  3. Fischler, Martin A.; Bolles, Robert C. (1981). "Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography". Communications of the ACM 24 (6): 381-395. doi:10.1145/358669.358692