[Reference] Prior, visualized

January 14, 2015

Input tree. (Mean of the local prior)

Training tree.

Sample from the locality prior only:

Samples from full prior (including epipolar constraints)

Notice how all corresponding points lie on the same epipolar line.

The plot of eigenvalues of the prior covariance matrix suggest a very low-dimensional embedding:

Plotting log-sqrt eigenvalues compresses the dynamic range, making it easier to read:

Posted by Kyle Simek
blog comments powered by Disqus