When generating the NV paths using GenerateNEOAngle, I want to simulate the resistance force of the background cell. For example, where there are vessels on the background image, then the underlying pixels have lower probability of being chosen, indicating that the resistance force of this pixel is high.

A optimal way to simulated the background force is to extract the background image beneath of the simulated NV and then assign some probabilities to each pixel based on the presence/absence of normal vessel.(normal vessel will hinder the growth of NV??)

Because the background image can only be obtained after the size of simulated NV is determined, so the aboved method can not be realized. So we turned to randomly assign some probabilies to the lattice sites of simulated NV.

On the meeting today, Dr.Li suggests me to validate the PLS method by visualizing the result.

It can be done in the following steps:

1 Use training data to get the PLS coefficients

2 Predict leaf variables using the PLS coefficients and the test data right hand variables

3 Plot the error between predicted leaf and true leaf

4 Visualize the human walking squence using the true right hand angles and predicted left hand angles.

Tom will do this starting June 7. I have to verify the result after he finishes.

I have done some experiments using the PLS to estimate the relationship between right
hand side angles and left hand side angles.

From all the experiments, I observed that the PLS2_25D_RegAngles (regress on the angles
but do Kalman Prediction as before) performs best.

The next step is to

1)check other regression methods like MLR,PCR. What percent of variance can they capture?

if higher then PLS, then use them for tracking, then compare the result.

2)Inter-comparison:

    PLS_RBPF_Tracking V.S. Annealled PF

    PLS_RBPF-Tracking V.S.RBPF tracking using curve fitting

3)Test PLS_RBPF_Tracking on new sequences such as running sequence which must have training and test data available

4)what happened if the nuber of regression samples is decreased or increased in the PLS regression?

5)HOw to compare with Annealled PF? using what kind of criterion as the error measure?

06/05/2006

Woking on revising the GenerateNEOAngle algorithm, such that

a)  it can generate NV branches in multiple levels and multiple branches at each level.

Purpose to increase the complexity of the NV appearance

b)add a resistance force to simulate the resistance force that comes from the background retina cell.