Jump to content

Pro eye unsuitable for research - data SUPER heavily filtered during fixation.


Recommended Posts

Any news on this topic? It's been one week, which is a long time in the VR world!

I'd love to buy a Vive Pro Eye today, but this makes me hesitate. In my research as neuroscientist/neuroengineer, the most exciting effects are real-time (foveated rendering, prosthetics, gaze-contingency), which depend on low latency. These effects will tolerate low accuracy/precision better than high latency.

So, does it look like there is a delay in registering eye movements relative to head movements? If the delay is constant, this is workable. If the delay is variable/unpredictable then the situation becomes more difficult.

 

Link to comment
Share on other sites

 

Here are the latency measurements taken by one of our students.  Images reflect the azimuthal component of the head and eye during vestibulo-ocular reflex (while fixating a stable object and rotating one's head about the vertical axis).  The first figure shows the original signals - deg/s over time.  The second shows the results of the cross correlation, and the third shows a zoomed in look at the signals before/after adjustment for the measured latency.  

The two signals were maximally correlated with 83 ms of shift, meaning that there was 83 ms of latency in the eye tracker relative to the head tracker. Vive head tracking is typically 22-33 ms absolute latency, so that puts the ET above 100 ms of absolute latency.  
 
We have only measured this once.  Really, we should measure a few more times, with different people.  We could also learn more by testing for possible effects of rotational velocity, and GPU load on the amount of latency.  
 
An effect of the former (head velocity) would indicate that this latency results form a dynamic filter.  This seems unlikely, because if it were, they would just have us turn it off.  
 
It's more likely attributed to the core functions of the eye tracking pipeline, such as pupil segmentation in the 2D imagery, or gaze mapping (which converts from 2D pupil centroids to 3D gaze vectors). If so, I would expect the latency to increase with GPU load.
 
Something to test, anyhow. Or, HTC could jump in and say a bit more...

image (1).png

image (2).png

image.png

  • Like 1
Link to comment
Share on other sites

Thanks gazeBehavior, this information is essential!

It sounds like the gaze delay is comparable to other current consumer VR eye tracking solutions around the same price point. It also sounds like the delay is due to software (gaze position calculation) rather than hardware.

To say that the Pro Eye eye tracker is unsuitable for research is too strong, as the eye data can be recorded and corrected after-the-fact. Additionally, task/gameplay demands could enforce longer fixations for real-time user control applications.

With a latency of 83ms relative to head tracking, the Pro Eye eye tracker is on the cusp of tolerable (50-70ms) although not optimal (20-40ms) latency for foveated rendering (https://dl.acm.org/citation.cfm?id=3127589).

This latency seems better than expected (although not as good as hoped). Gaze position calculation could be optimized for speed; however, optimization may require alterations in tobbi's proprietary (and encrypted) code, and therefore might be outside of HTC's and developers' control.

Edited by VrHacker
Link to comment
Share on other sites

The SMI eye tracker was far below this latency before Apple took it off the market, but what can we do about that?  It also cost a LOT more when it was available. 

Quote

To say that the Pro Eye eye tracker is unsuitable for research is too strong.

  I agree.  I wrote that when I was led to believe (by an HTC dev's comment) that this latency was actually the result of a dynamic filter.  It wasn't until I turned the filter off that I settled on the interpretation that the latency was due to core algorithms, and not a dynamic filter.

Quote

...eye data can be recorded and corrected after-the-fact

I also agree that, if it the latency is stable within a session and across sessions, we might be able to adjust for post-hoc, offline data analysis.  However, we do not yet know how stable the latency is across a single session, and across multiple sessions. One could also measure latency using the VOR trick - it is fairly quick to record.  

Quote

With a latency of 83ms relative to head tracking, the Pro Eye eye tracker is on the cusp of tolerable (50-70ms) although not optimal (20-40ms) latency for foveated rendering (https://dl.acm.org/citation.cfm?id=3127589).

I'll add that it's possible that the 80 ms I saw on my machine could be brought down on a very new machine (mine was great when I bought it, but is now 3 years old).  I'm abroad, but can post the specs when I get back into the USA.  

@Daniel_Y @Cory_HTC @Jad

Link to comment
Share on other sites

  • 2 months later...

Create an account or sign in to comment

You need to be a member in order to leave a comment

Create an account

Sign up for a new account in our community. It's easy!

Register a new account

Sign in

Already have an account? Sign in here.

Sign In Now
 Share

×
×
  • Create New...