Re: Small change proposal foricvCalcOpticalFlowPyrLK_8uC1R

Adi,

I think, such a change will be useful.

Regarding possible optimizations. Because of the covariation matrix properties you do not need to compute both eigen values and choose min. It will be just:

double smallEig = (Gxx + Gyy – sqrt((Gxx – Gyy)*(Gxx – Gyy) + Gxy*Gxy)))*0.5;

Regards,

Vadim

Hi,

  For a project I've been working on, I needed to know the confidence of the LK tracker result.
Although the status vector gives a binary answer, it is not always enough.

One way to know if the tracked point is not on a corner but on an e.g. is by checking the eigenvals of the 2x2 derivative matrix.
A small eigenvalue means that the matrix is badly conditioned and that either this is an edge (a single small eigenval == aperture effect), or that this is a smooth featureless region (2 small eigen vals).

These situation can happen when the points are not from GoodFeaturesToTrack (which is heavy) or after tracking point over a long time period where drift can happen.

I made a small non-API breaking change to cvlkpyramid.cpp:

*** cvlkpyramid.new.cpp Wed Apr 30 14:10:31 2008
--- cvlkpyramid.org.cpp Thu Aug 31 19:14:36 2006
***************
*** 541,561 ****
                      }

                      D = Gxx * Gyy - Gxy * Gxy;
-
                      if( D < DBL_EPSILON )
                      {
                          pt_status = 0;
                          break;
                      }
-
-                     //////////////////////////////////////////////////////////////////////////
-                     // Adi Shavit - 2008.04
-                     double smallEig  = MIN( -(sqrt(Gyy*Gyy-2*Gxx*Gyy+4*Gxy*Gxy+Gxx*Gxx)-Gyy-Gxx)/2,
-                                              (sqrt(Gyy*Gyy-2*Gxx*Gyy+4*Gxy*Gxy+Gxx*Gxx)+Gyy+Gxx)/2);
-
-                     pt_status = MAX(1,cvRound(smallEig/(jsz.height*jsz.width)));
-                     //////////////////////////////////////////////////////////////////////////
-
                      D = 1. / D;

                      prev_minJ = minJ;
--- 541,551 ----
***************
*** 582,594 ****
              }

              featuresB[i] = v;
!
!             //////////////////////////////////////////////////////////////////////////
!             // Adi Shavit - 2008.04
!             status[i] = (char)MIN(pt_status,255);
!             // status[i] = (char)pt_status; // original line
!             //////////////////////////////////////////////////////////////////////////
!
              if( l == 0 && error && pt_status )
              {
                  /* calc error */
--- 572,578 ----
              }

              featuresB[i] = v;
!             status[i] = (char)pt_status;
              if( l == 0 && error && pt_status )
              {
                  /* calc error */

It calculates the closed form solution for the 2x2 eigenvalues, when the Det != 0, and chooses the smaller value.
It then normalizes the eigval by the window size.
This allows the value to fit relatively comfortably in the status var which is (unfortunately) of type char.

The status char now returns 0 for bad tracking and non-zero value when the tracing succeeds as before.
However, the actual value of status is the normalized smallest eigenval.

The code can be a bit more optimized, but I'd like to know if this kind of change would be acceptible here?

What do you think?

Thanks,
Adi