#
[Work Log]

June 26, 2013

## Tasks

- Likelihood construction,
`clean_correspondence3.m`

- new tracking GP model

## Likelihood construction

Started and finished implementation today. Need to design a test and then debug.

~ 250 lines of Matlab code. Logic overview:

- smooth triangulated points
- resample smoothed 3D curve uniformly
- handle tails (see below)
- project curve into each view and resample uniformly
- DTW to correspond 2d data curve to projected smooth curve (see below)
- map corresponding projected curve points back to 3D points and indices
- triangulate 2d data points against corresponding 3d point
- compute likelihood hessian around that point

## New DTW

Re-implemented a specialized version of DTW with following changes:

- only horizontal steps accrue cost
- Hard-constraint on the number of vertical steps per horizontal step.
- keeps track of "best" match along vertical runs. no need for second pass
- I think this is only possible because only horizontal steps accrue cost.

## Tail points

Still iffy on the "tail points" case. Implemented late at night and likely needs review in the morning. Still need to handle negative index values.

Alternative implementation: only inspect the tails to determine the length of the 3D curve. Then proceeed as usual. no special cases

## TODO

- review tail points code
- consider alternative implementation (see above)

- handle negative index values
- think about hessian and transformation Jacobian
- testing, debugging, profiling

Posted by
Kyle Simek

blog comments powered by