10/3/2003

Rewrote the occlusion handling routines. It appears better:

3+1 Sequence (I also called it "Problem Sequence" earlier)
Union3 Sequence
Simple Overhead Sequence (I say simple because the movement is too smooth)

Here is another sequence.
At the end of summer I had a simple occlusion "handling" method working.
I've run the current program for comparison:
Previous
Current

In some cases the current routine looks like it did a worse job.
For example, in the JTcross sequence it loses the right-most person
more than previously. However, it generally looks slightly cleaner.
And in the harder Union3 sequence it doesn't lose as many people.

9/26/2003

New and improved results this week:

Problem Fixed
Less Confusion
Longer Sequence
New Sequence

The latest incarnation of the program solves a problem where an object
would jump from one person to another who didn't even look like the first.
Well...maybe SOLVES is not the mot juste, but it definitely ameliorates the situation.

* The original criteria for grouping patches by vector is to find the largest group.

* Criteria modified to balance number of patches and correlation fitness.

I.E. Group with highest total correlation is chosen.

9/19/2003

Here are links to some sequences. The first was processed under an old version of the program and is included to show progress. The second is the improved version. The third is the background subtraction for the first two. The last shows some more problems with the program.

Sequences:

Union3 Reference for comparison.
Union3 Improved Less confusion (a little less) during partial occlusion.
Tracking also more accurate during partial occlusion.
Background Subtraction for Union3
Problem Sequence Doesn't understand when left person disappears.