Unless otherwise noted, all filesystem paths are relative to the "Working path" named above.

Continuing: ground-truth-to-data-labels. See file train/label_from_ground_truth.m.

Finished.

Next:

Data labels to likelihood means/covariances

Likelihood means/covariances to marginal likelihood

training framework

training

Visualizing labels_from_ground_truth()

Wrote tmp_get_max_posterior.m, a temporary script that computes the posterior mean from a possibly-overconstrained prior. In this case, the posterior covariance is singular, but the mean can still be obtained. The math behind it is available here.

Below is a plot using the test dataset and the ground truth labels:

This shows that given a good labeling, a quality 3D reconstruction can be obtained using only the fragmented curves output by the curve-detector.

Notice that the curves at the base have missing parts. There isn't sufficient edge data here, but this could probably be fixed by connecting them to the base of the main stem and using the Branching Gaussian Process prior to enforce connectivity.