1 | function y = compare_real_motion(fo,D,s,n) |
---|
2 | % This is meant to be like compare_motion, but with a real motion matrix |
---|
3 | % instead of generating a random one. |
---|
4 | num_random_starts = 5; |
---|
5 | nframes = size(D,1)/2; |
---|
6 | npoints = size(D,2); |
---|
7 | |
---|
8 | for j = 1:n |
---|
9 | INC = motion_incidence(fo,nframes,npoints); |
---|
10 | Dtj = remove_translations(D, INC); |
---|
11 | M = approx_full_matrix(Dtj,3); |
---|
12 | ERR = Dtj - M; |
---|
13 | if rem(j,5)==0 |
---|
14 | y = [y', compare(M,INC,3,ERR,s,num_random_starts)']'; |
---|
15 | else |
---|
16 | y = [y', compare(M,INC,3,ERR,s,num_random_starts)']'; |
---|
17 | end |
---|
18 | end |
---|
19 | |
---|
20 | % First of all, we remove translation as best we can from the occluded data, |
---|
21 | % and then give all the methods the same translation-free problem. This |
---|
22 | % should provide a level playing field, while indicating the noise that |
---|
23 | % occurs when there might be errors in determining translation. The other |
---|
24 | % alternative would be to give ground truth a translation based on all the |
---|
25 | % data. |
---|
26 | |
---|
27 | % Now, M is our pseudo-ground truth; ie. the best guess to the error free matrix. |
---|
28 | % M + ERR is Dtj, which is the data matrix with translation removed, as best |
---|
29 | % we can. The "error" in ground truth will be the difference between our |
---|
30 | % best guess of M and the data; sounds ok. |
---|
31 | % If we feed M and ERR to 'compare', we'll be ok since M is only used: |
---|
32 | % a) added to ERR to give the data matrix. This is in fact what we're trying |
---|
33 | % to approximate with a rank three matrix. and b) as the ground truth. |
---|