[Work Log] Misc.

Monday IVILAB infrastructure meeting

I'll be organizing the Computational Intelligence seminar this semester.

Will need to

• Arrange volunteers
• send announcements
• send reminders to speakers
• Set up seminar web page

Building efficient curve-tree ML.

Significant rework.

New field: branch_K and branch_mu to store mean and covariance of all points in curve. Prior_k now does not include offset covariance; stored in branch_K. mu_b and Sigma_b now store the branch point means and covariances for all views

fix kernel to not include offset index

Issues

Finished implementation. Has bugs.

Start-point has no effect

Example:

start_pt = [0 100];
ml = [];
for i = 1:2
    Corrs = attach(Corrs, 2, 1, start_pt(i), 0, params);
    ml(i) = curve_tree_ml_2(Corrs, params, data_);
end
assert(ml(1) == ml(2));

Fixed - stupid bug in attachment/attach.m - was index offset was hard-coded to zero due to a refactoring mishap.

Optimal Branch point isn't correct

Example:

start_pt = [0 10 50 100 1000 5000];
ml = [];
for i = 1:2
    Corrs = attach(Corrs, 2, 1, start_pt(i), 0, params);
    ml(i) = curve_tree_ml_2(Corrs, params, data_);
end
assert(all(diff(ml) > 0);

Solved - curves were reversed.

Handling reversed curves

Should add a reversed flag, which reverses indices before building likelihood and prior. currently, constructing the likelihood occurs during the "backproject and re-index" phase, in file correspondence/corr_to_likelihood.m. Should refactor likelihood construction into its own function, so we don't have to re-backproject when we don't need to.

Let's review the data-flow so we can see more clearly where everything happens.

Overview: end-to-end curve construction

1. propose association and attachment
   • By sampling (no code yet)
   • train/labels_from_ground_truth - propose from ground truth.
   • Output: assoc {}
2. Construct track
   • correspondence/make_correspondence
   • Output: Corrs {}; Corr.ml_2d
3. Correspondence and triangulation
   • correspondence/build_full_correspondence.m - build from scratch
   • correspondence/merge_correspondence_2.m - merge two pieces
   • Output: corr, means, precisions, cov_error
4. backproject and estimate curvature
   • correspondence/corr_to_likelihood.m -
   • Output: ll_{means, precisions, indices}, curve_sm*
5. construct likelihood (flatten, sort, and reverse if needed)
   • correspondence/flatten_sort_and_reverse.m - (doesn't exist yet)
   • Output: ll_{means_flat, precisions_flat, indices_flat, S}
6. handle attachment recursively
   1. compute conditional prior
      • attachment/att_set_start_index.m
      • Output: start_index, prior_K, prior_indices
   2. compute branch point posterior
      • attachment/att_set_branch_index.m
      • Output: branch_index, mu_b, sigma_b, branch_K, branch_mu
7. Compute (ML, argmax, etc)

Consider renaming step 4. backproject against rough triangulation; estimate curvature at each point; determine index set. At this point, the order of points don't matter, because the index set hasn't been put to use.

Action: reverse curve - detach?, rerun step 5, 6.1 & 6.2 on self, update branch point & rerun 6.2 for children.

Test: optimize branch point and start index

(TODO)

TODO

• optimize test
• get attachment, reversal, and branch points from ground truth
• store ml_2d with corr. update on merge. use during ml computation instead of data_

Posted by Kyle Simek

blog comments powered by Disqus