Example Results of "HALF-SIFT" (CVPR09)

---

HALF-SIFT uses the SIFT [1] framework for detecting scale invariant feature points in an image and establishing correspondences between images. By replacing the subpel interpolation scheme of the original SIFT by a regression with an image signal adapted function, the accuracy of the localization of detected points is increased by up to 16% for natural image pairs. Two proposals are made:

• Gaussian regression function
• Difference of Gaussians (DoG) regression function

Results (error histograms of detected corresponding feature points) are shown below

• input data: image pairs with constant extrinsic parameters, varying illumination ['cars', 'towers', 'gift']; jpeg-artefacts ['ubc'] .
• see paper "HALF-SIFT: High-Accurate Localized Features for SIFT" [2] for details; 'cars' and 'ubc' seq included in the paper

Image Pairs:

Images	SIFT vs. HALF-SIFT - DoG	SIFT vs. HALF-SIFT - Gaussian	Download tif images (1920x1080)
cars A: constant_intrinsics
cars B: auto-iris
cars C: automatic_gain
towers A: constant_intrinsics
towers B: auto-iris
towers C: automatic_gain
gift A: constant_intrinsics
gift B: auto-iris
gift C: automatic_gain
ubc: img1 - img2
ubc: img1 - img3
ubc: img1 - img4

References

• [1] David G. Lowe.
  "Distinctive image features from scale-invariant keypoints"
  International Journal of Computer Vision (IJCV),
  60(2):91-110, 2004.
• [2] Kai Cordes, Oliver Müller, Bodo Rosenhahn, Jörn Ostermann
  "HALF-SIFT: High-Accurate Localized Features for SIFT"
  IEEE Workshop on Feature Detectors and Descriptors: The State Of The Art and Beyond, in conjunction with IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2009, pdf