hello everyone
As we can see, moving the head in depth greatly affects gaze Y estimation (up-down)
That is, I believe, because Y difference between Pupil and Glints changes during head movement is Z, and tracker wrongly thinks that I looked up or down
Before Martin, Javier and team offer Y-error compensation in gaze estimation algorythm (hope soon),
I suggest trying to compensate Y error using measures provided by API currently, I mean pcenter Left and pcenter right.
The idea is as follows:
- calculate the average distance between pcenter left and pcenter right during calibration
- adjust Y gaze coordinate streamed by API after calibration using the formula like (initial Y provided by API)/(average distance between pupils during calibration)*(observed distance between pupils)
(the coefficient must be different for different screen areas, some filter should be applied, etc., but simple formula looks enough for rough correction of big error)
what do you think? who can try (ny experience in programming is not much better than just copy-paste)?
It is much better to include such correction into calibration and estimation process, but i dont know neither C# nor math unfortunately (and the code is not open too)
I included such correction previously in open-source ITU gaze tracker code (I did it rather intuitively due to lack of knowledge) and it worked perfectly (in terms of Y-error compensation), but that time it became much harder to finish calibration, I dont know why
update below
The error of head movement is very big, especially for me (due to glasses I have to place tracker closer to face), and both in Y and X (try it at screen corners), though X-error is much smaller
I implemented a very rough correction in TETWinSamples code, C#
Start with Calibration, then press Mouse Control, please find below (i'm not a programmer neither a mathematics-skilled, make it better if possible)
The last portion of code is formula that can be applied to coordinates reported by API
TETWinSamples code is open-source one, and I hope I did not brake any rules, if yes - I apologies and ask moderator to delete the post
below the changes that greatly decrease the error for my setup If something does not work, I can send files, but can not help with programming at all
1.Make set of global variables necessary for correction to connect different parts of code together. I inserted the code into TrackBoxStatus.xaml.cs, Line 22, right after
namespace TETControls.TrackBox. Not the best place and solution I guess, but it works
- Code:
class MyCorrection
{
public static double [] ave_dist_arr = new double [99999];//!!! Array of distances between pupils during calibration, 99999 is a max number of frames for calculation (may be much lower for sure)
public static double ave_dist;//!!! Average distance between pupils during one frame
public static int calib_samples_count;//!!! Count of frames in array
}
2.Fill the array of distances between pupils during calibration. I placed an addition to TrackBoxStatus.xaml.cs, Line 145 (approximately), right inside
public void OnGazeUpdate(GazeData gazeData)
Adding a condition on “calibration is processing” would be good but not necessary. The array should be filled in only during calibration, but I didnt find the right place due to absence of programming knowledge
- Code:
//!!! Filling the array
if (TETControls.TrackBox.MyCorrection.calib_samples_count < TETControls.TrackBox.MyCorrection.ave_dist_arr.Length)
{
TETControls.TrackBox.MyCorrection.calib_samples_count += 1;
TETControls.TrackBox.MyCorrection.ave_dist_arr[TETControls.TrackBox.MyCorrection.calib_samples_count] = Math.Sqrt(Math.Pow((gazeData.RightEye.PupilCenterCoordinates.X - gazeData.LeftEye.PupilCenterCoordinates.X), 2) + Math.Pow((gazeData.RightEye.PupilCenterCoordinates.Y - gazeData.LeftEye.PupilCenterCoordinates.Y), 2));
}//!!!
3.Renew all variables when calibration is started. I placed the code into CalibrationRunner.xaml.cs, Line 213 (approximately), right before DoStart();
- Code:
//!!! Clear variables
Array.Clear (TETControls.TrackBox.MyCorrection.ave_dist_arr, 0, 99999);//!!!
TETControls.TrackBox.MyCorrection.calib_samples_count = 0;//!!!
//!!!
4.Calculate the average distance between pupils during calibration based on array filled. I placed the code into CalibrationRunner.xaml.cs, Line 375 (approximately), right before StopAndClose();
- Code:
//!!! Calculate average distance
double sum = 0;
int elem=0;
for (int i = 0; i < TETControls.TrackBox.MyCorrection.ave_dist_arr.Length; i++)
{
if (TETControls.TrackBox.MyCorrection.ave_dist_arr[i]>0)
{
sum = sum + TETControls.TrackBox.MyCorrection.ave_dist_arr[i];
elem = elem + 1;
}
}
TETControls.TrackBox.MyCorrection.ave_dist = elem != 0 ? (double)sum / elem : 0;
Console.WriteLine("Done, average distance = " + TETControls.TrackBox.MyCorrection.ave_dist);//!!!
///!!!
5.Use calculated average distance to correct error in Mouse Control. Replace screenX and screenY formula in CursorControl.cs, Line 59 (approximately), with corrected one, like below. I know the formula is not very good, but the result is much better than one without correction. Please help with better correction formula. The one I used is not very good and does not depend on screen area (but it should I guess). Anyway, in case distance between pupils does not change significantly, the correction does not correct coordinates significantly too
- Code:
//var screenY = (int)Math.Round(y + gY, 0); // Original formula
//var screenX = (int)Math.Round(x + gX, 0); // Original formula
//!!! Correction
double pup_dist = Math.Sqrt(Math.Pow((gazeData.RightEye.PupilCenterCoordinates.X - gazeData.LeftEye.PupilCenterCoordinates.X), 2) + Math.Pow((gazeData.RightEye.PupilCenterCoordinates.Y - gazeData.LeftEye.PupilCenterCoordinates.Y), 2)); // pup_dist - current distance between pupils. If differs from the one during calibration, than the tracker calculates both X and Y wrongly
double ave_d = TETControls.TrackBox.MyCorrection.ave_dist;
if (ave_d == 0) ave_d = pup_dist; // if not calibrated during the session, ave_d = 0, correction would not work
var screenY = (int)Math.Round(gazeData.RawCoordinates.Y - (gazeData.RawCoordinates.Y - ActiveScreen.Bounds.Height*1.25)*((pup_dist-ave_d)/ave_d)); // I would very appreciate if you advise better formula for Y. 1.25 means 1 (screenHeight) + 0.25 (distance from screen bottom to device)
var screenX = (int)Math.Round(gazeData.RawCoordinates.X - (gazeData.RawCoordinates.X - ActiveScreen.Bounds.Width / 2) * ((pup_dist - ave_d) / ave_d)); // I would very appreciate if you advise better formula for X
//!!!
Statistics: Posted by skatak.aizenberg@gmail.com — 22 Apr 2015, 12:54
]]>