[Work Log] Work Log
Implemented nonmaximal suppression, background subtraction.
Derived/implemented expression for cubic covariance function for negative time-points.
Tested/debugged branch parameter estimation. Implemented a visualization that revealed several issues during optimization. Found a bug in gradient, which improved performance. Also realized that adding a constant perturbation per stem allows much better fitting for touching/crossing curves. To implement this, I created a "hamming distance" covariance function, which adds a constant offset to each child. Also, increasing the value on the diagonal improves gradient estimates, making optimization converge faster.
Incorporated parent estimation into end-to-end fitting routine.
Next:
Post-tracking tasks:
- triangulate, smooth, etc
- shrink/grow stems in 3d
- implement 3rd-view fitting
- try adding new stems ** given current fit, run 2d stem finding
General procedure for fitting:
- find a stem in 2D
- estimate a perturbation in 2D
- triangulate and fit in 3rd view (w/ epipolar constraint)
- attempt to infer parent
- accept/reject
Open issues
- width fitting
- nonconstant width?
- fit independent curves independently?