A Concurrent Comparison of EyeLink 1000+ and TrackPixx3 on a Benchmark Test Battery

Anna Prysłopska

anna.pryslopska@ling.uni-stuttgart.de

University of Stuttgart
Institute of Linguistics

Benedikt Ehinger

benedikt.ehinger@vis.uni-stuttgart.de

University of Stuttgart
Institute for Visualization and Interactive Systems

Titus von der Malsburg

titus.von-der-malsburg@ling.uni-stuttgart.de

University of Stuttgart
Institute of Linguistics

EyeLink 1000+

gold standard

SR Research

TRACKPixx3

recent debutante
VPixx

The strengths and weaknesses of EyeLink 1000+ are well-understood. There is no comprehensive evaluation of TRACKPixx3.

	EyeLink 1000+	TRACKPixx3
Recording	2kHz binocular	2kHz binocular
Range (desktop)	40-70 cm	60-160 cm
Latency	1.3-1.8 ms	1.7 ms
Output	EDF	user defined
Chinrest	optional	required
Price	negotiable	10-20% cheaper
Publications	~ 3360	~ 18

range depends on mount

EyeLink 1000+ vs. TRACKPixx3

concurrently while recording eye movements

• 7 tasks × 6 blocks
  Ehinger, Groß et al (2019)
• decay of accuracy
• fixation durations
• pupil dilation
• smooth pursuit movement
• microsaccade classification
• blink classification

explain setup, one illuminator (TPX due to better results), chin rest used

Tasks

Data analysis

We used the Robust Eye Movement Detection for Natural Viewing (REMoDNaV) for event classification.

other event classification algorithms may yield different results

TrackPixx data is filtered by a proprietary non-linear filter before export.

For a fairer comparison, we filtered EyeLink and TrackPixx data using a Savitzky-Golay filter.

this means raw trackpixx = native filter, trackpixx filtered = filtered twice (native + us) filter length 62 ms and polyorder 3 We iteratively tested various parameters until we reached a satisfactory result. The filtering includes a Gaussian spatial-temporal filter. It also implements non-linear (eg: median filtering) spatio-temporal image processing components as well. In addition, it has very low-pass filtering of certain camera characteristics, to compensate for temperature/voltage drift of the camera’s analog-to-digital converters.

Grid viewing

The participants were instructed to fixate targets that appeared at the crossing points of a 7 × 7 grid. The participants were asked to saccade to the target and fixate it, and once they felt their eyes stopped moving, to press the space bar to continue. Start and end in the middle. 49 targets were shown during every task repetition.

Accuracy was comparable between both eyetrackers. TrackPixx had higher precision before (re-)filtering.

The recorded fixations were the same length.

Free viewing

For the free viewing task, we presented 18 different photos of natural images. Participants fixated on a central fixation cross. The participants were instructed to freely explore the images. During each of the six blocks, we showed three images for 6 s.

EyeLink fixations & saccades

Trackpixx fixations & saccades

The eyetrackers had similar precision.

these are not filtered by us. you can tell how smooth trackpixx is

EyeLink raw and filtered,
TrackPixx original and filtered.

EyeLink recorded longer fixations but the differences mostly disappeared after filtering.

EyeLink raw and filtered,
TrackPixx original and filtered.

TrackPixx recorded shorter saccades, but the differences diminished after filtering.

Both eyetrackers recorded the same amount of fixations.

before and after filtering, so this does not seem to be affected

Microsaccades

we showed a central fixation target for 20 s. The participants were instructed to continue the fixation until the target disappeared.

TrackPixx recorded more saccades before filtering.

but then the numbers even out

The saccade amplitude was larger for raw EyeLink.

but then the numbers even out

Compared to TrackPixx, raw EyeLink recorded higher peak velocity, and fewer and larger microsaccades. EyeLink detected biggest and fastest microsaccades, while TrackPixx detected the slower ones.

when line is straight then it’s all ok, bc linear relationshisp between peak velocity and amplitude, but when it’s not then something strange is happening; microsaccades; what to say: peak velocity in EL unfiltered is higher than microsaccades were fewer but bigger for eyelink eyelink only detected fast biggest microsaccades, but other one only got the slow ones

again, no difference after filtering, but information may suffer

Blink detection

The participants fixated a central fixation target and were instructed to blink each time they heard a beep. The beep sound chimed 100 ms for seven times with a pause of 1.5 s between every beep.

Both eyetrackers recorded the same number of blinks. TrackPixx recorded longer blink durations.

Pupil dilation

In this task, we varied the light intensity of the monitor to stimulate a change of pupil size. During the entire task, a central fixation target was displayed which the participants were instructed to fixate. Each block consisted of four different monitor luminances 25%, 50%, 75%, and 100%. Before each target luminance, we first showed 7 s of black luminance.

The eyetrackers recorded the similar pupil area, with TrackPixx recording somewhat larger values.

each pupil area value (pa) is divided by the median pupil area value during the baseline period

Reading

Both eyetrackers recorded the same number of fixations within an area of interest.

first a fixation point where the first word appeared. then the Tulip Mania text in each block just a few sentences, whole story throughout the experiment. Click a button when done. Summarize at the end.

Participants read the “Tulips” story from the Natural Stories Corpus (Futrell et al. 2021) story of the tulip mania. The original text is in English but we translated it into German.

Smooth pursuit

see upcoming paper

with a target moving from the center of the screen outward using 24 different angles and five different speeds

Conclusion

The eyetrackers performed similarly overall.

TrackPixx’s unknown native filter may affect PSO and microsaccade detection.

EyeLink 1000+ continues to be the benchmark.

TRACKPixx3 presents itself as a viable alternative.

Thank you