DKM: Dense Kernelized Feature Matching for Geometry Estimation

Johan Edstedt, Ioannis Athanasiadis, Mårten Wadenbäck, Michael Felsberg
Computer Vision Laboratory, Linköping University   
Paper Code
Visualization of DKM warps.
We visualize the warp between the images by sampling the estimated matching pixel in the other image.
We multiply this value with the estimated confidence.
The animation is done by interpolating between the original image and the warped one.

TLDR

Instead of finding sparse correspondences we estimate a dense warp.
In contrast to previous dense matching work, our dense approach is able to achieve state-of-art results on two-view geometry.

Method

We propose Dense Kernelized Feature Matching (DKM). Our approach consists of three main contributions:

  1. A Gaussian Process based Global Matcher.
  2. Large depthwise kernel warp refinement.
  3. A balanced sampling approach to estimation.

BibTeX


      @inproceedings{edstedt2023dkm,
        title={{DKM}: Dense Kernelized Feature Matching for Geometry Estimation},
        author={Edstedt, Johan and Athanasiadis, Ioannis and Wadenbäck, Mårten and Felsberg, Michael},
        booktitle={IEEE Conference on Computer Vision and Pattern Recognition},
        year={2023}
      }