Improving Eye Tracking Accuracy

[Tuxie]

Hi,

 

I'm using the VIVE Pro Eye for a kind of eye tracking research at a university. Thank the VIVE team for developing the impressive device. But I'm facing some difficulties about the product. Could someone give me advice?

 

I'm modifying the code of SRanipal_example and sample_vs2015 of SRworks with c++ environment.

 

A. The gaze points from both eyes should be at the same point when the player looks at approximately infinity far location (like 100 m far) with see-through camera. But the data ('gaze_direction_normalized') from SRanipal are not. Why?

 

B. To correct the above problem, I think the IPD calibration can be done as follows. Is it correct or not recommended?

    1. The player gazes at approximately infinite far (like 100 m) far through the see-through function.

    2. Adjust the knob so that  'gaze_direction_normalized' of both eyes (0,0,1) (or the similar value)

 

C. In my environment, the right eye gaze point is usually at lower point than the left eye (please see images). Do you have any idea about the cause?

 

D. Eye tracking data is not consistent. The disparity between eyes occasionally becomes larger or smaller.  Can I improve it?

 

 

   Tag added by moderator

[Daniel_Y]

1. Yes, you need to lauch  eye tracking calibration from VR dashboard first.

2. refering this thread for coodinate explaination, https://community.viveport.com/t5/Vive-SRanipal-SDK/Vive-Pro-Eye-Finding-a-single-eye-origin-in-world-space/m-p/31592#M10

3. Refering to eye tracking accuracy in VIVE PRO EYE SPECS, https://www.vive.com/eu/product/vive-pro-eye/

4. Combined gaze may be more fit your use scenario.

 

[Tuxie]

Thank you so much for the rapid reply.

 

1. Sure. Even after the calibration, both eye fixations don't converge at the virtually infinite far object.

2. I knew the thread. Based on the geometry, the infinite far object is drawn at the same point in each display, right? Or the display of the vive is not alligned at the center?

3. I think this phenomenon is not Gaussian distribution...

4. Do you mean "convergence_distance_mm"? If so, yes it'll help me much. But it is not implemented yet, right?