Unconstrained Image Matching

This research focuses on automatically determining dense and smooth mapping between two images without a priori knowledge of either the camera pose or objects. In order to find plausible correspondences, images are first convolved with a set of linear filters similar to those that are used in early vision. The vector fields consisting of filter responses are then matched with each other by minimizing the cost function which expresses the similarity of transformed images and mapping smoothness in a multiresolutional hierarchy. Since a point in one image can correspond to its counterpart far away, globally optimized mapping may result in strong local distortion. Such distortions are prevented by guaranteeing the convexity of mapping in nonlinear optimization.

Publications

Shuntaro Yamazaki, Katsushi Ikeuchi and Yoshihisa Shingawa,
“ Plausible image matching: determining dense and smooth mapping between images without a priori knowledge ”,
International Journal of Pattern Recognition and Artificial Intelligence, vol.19, num.4, pp.565-583,
June 2005

Shuntaro Yamazaki, Katsushi Ikeuchi and Yoshihisa Shingawa,
“Determining plausible mapping between images without a priori knowledge”,
In Proc. Asian Conference on Computer Vision 2004 (ACCV2004), vol.1, pp.408-413,
Jan 2004

Shuntaro Yamazaki, Katsushi Ikeuchi and Yoshihisa Shinagawa,
“Unconstrained view-interpolation based on automatic image-morphing”,
In Proc. Meeting on Image Recognition and Understanding 2004, vol.2, pp.222-227,
July 2004
(Japanese)

Supplementary Material

Demo video

This video is a compilation of the main results of this project. (also uploaded to YouTube.com)

Figures in Paper

We used 24 filters composed of first and second partial derivatives of Gaussian, with 4 scales, in 2 and 3 orientations for the first and second derivatives respectively. The filters have zero mean and each of them is divided by its L1 norm for scale invariance.

The convexity of mapping is maintained as long as each point m(px,y) in a source image (a white point on the left) is mapped into a quadrilateral as illustrated by a hutched area on the right.

Robustness: Since mean filter and CPF are directly affected by the change of image intensity, resulting mappings are totally disrupted, while our method yields a plausible mapping.

Mapping determined by our algorithm.

Results

Registration of MRI images.