Performance of Wide-baseline Matching Using DAISY

Dennis Grootendorst

Department of Information and Computing Sciences, Utrecht University.

Abstract

This webpage is about an experimentation project performed by Dennis Grootendorst under supervision of dr. Robby T. Tan and dr. ir. Nico P. van der Aa. The goal of this project is perform extensive experimentation on DAISY to determine its matching performance. The DAISY descriptor is a local region descriptor used to obtain computational efficiency for dense matching. We show how DAISY obtains its computational efficiency, but builds similar histograms as the Scale-invariant feature transform (SIFT) descriptor. SIFT is a local region descriptor well-known for its performance on matching. The performance of the DAISY descriptor on wide-baseline (stereo) matching is analyzed and more insight is given in how the parameters should be to give the best results.

Paper

Performance of Wide-baseline Matching Using DAISY [PDF]
Dennis Grootendorst

Results

We have set out the robustness of the DAISY descriptor against the baseline. The figure below shows the result of this experiment: the robustness decreases linearly as the baseline increases.

DAISY provides three different normalization methods. By default DAISY uses the `Partial` normalization method which should give more accurate results near occlusions. Occlusions are larger at a wider baseline. Our results showed that the Partial normalization method performs worse at any baseline. See the figure below.

We have implemented SIFT for computing a SIFT descriptor for a single pixel in an image. We made a comparison between the original SIFT code of Rob Hess' on which our code is based. The descriptor values of our implementation and Rob Hess' implementation are set out in the figure below. For the implementation see the section Supplementary Materials.

Click for a larger image.

Supplementary Materials

Our system for comparing SIFT with DAISY uses an implementation of SIFT based on the SIFT code of Rob Hess. This implementation computes a SIFT descriptor for a single pixel in an image.

Implementation of SIFT for computing a descriptor for a single image.

References

Engin Tola, Vincent Lepetit, Pascal Fua. A fast local descriptor for dense matching. Proceedings of computer vision and pattern recognition (CVPR June 2008). Alaska, USA.
Rob Hess. An open-source SIFT Library. In Proceedings of the international conference on Multimedia (MM '10). ACM, New York, NY, USA, 1493-1496.