In this article we explain what eye tracker sampling frequency is and its impact on data collection.
The sampling frequency of an eye tracking system refers to how many times per second the position of the eyes is registered by the eye tracker. The higher the sampling frequency, the better your ability to estimate the true path of the eye when it moves. However, a higher sampling frequency also comes at a cost; more expensive cameras, more illumination needed, potentially higher noise levels, and ultimately more data to store. So, it makes sense to understand your own needs before deciding on what sampling frequency would be right for you. We hope by the end of this article you will know enough to make that decision.
Figure 1. Horizontal movement of the eyes where they appear to make a fixation-saccade-fixation sequence. The black line represents the true movement path of the eye over time. The vertical lines represent the different moments the eye tracker collects a picture of the eyes and estimates their position and gaze point (i.e. a sample).
Consider the movement in Figure 1 above. Visualizing only the horizontal movements, there appears to be a fixation, followed by a saccade, which is then followed by another fixation. This movement can be sampled by one eye tracker with a certain sampling frequency (blue), or it can be sampled by an eye tracker with twice that sampling frequency (red). You will also notice that the interval between each sample, the sampling interval, is shorter. If an eye tracker samples at 60 Hz, then the sampling interval is 16.67 milliseconds (ms), and if you double the sample frequency to 120 Hz you will also halve the sampling interval to 8.33 ms. If your eye tracker samples at 600 Hz, or even 1200 Hz, then the sampling interval is only 1.67 ms or 0.83 ms, respectively.
If you try to reconstruct the original waveform (the movement path) using these two sampling frequencies, you will not get completely similar results (see Figures 2 & 3, below). Judging just from the reconstructed waveforms, the blue, more sparsely sampled, waveform gives the impression that the saccade is a very sudden movement from one point to another. The other waveform reveals move nuance in the movement. This saccade is large enough to also be picked up by the slower sampling frequency, but that may not have been the case if it had been a very small saccade, or if we had used an eye tracker with a much slower sampling frequency.
Figure 2. The same horizontal movement of the eyes, as figure 1, with a fixation-saccade-fixation. This time the black line represents the estimated path of the movement of eye based on a low sampling frequency. The vertical lines represent the different moments the eye tracker collects a picture of the eyes and estimates their position and gaze point (i.e. a sample).
Figure 3. The same horizontal movement of the eyes, as figure 1, with a fixation-saccade-fixation. This time the black line represents the estimated path of the movement of the eye based on a high sampling frequency. The vertical lines represent the different moments the eye tracker collects a picture of the eyes and estimates their position and gaze point (i.e. a sample).
If you’re interested in studying the characteristics of the eye movements themselves, an eye tracker with a higher sampling frequency (and low noise levels) allows you to study microsaccades and other fixational eye movements, or other phenomena such as post-saccadic oscillations, or investigate saccades in detail (e.g., Juhola, Jäntti, & Pyykkö, 1985). For most researchers, the main benefit is a more robust transformation of raw gaze samples into fixations and saccades, with less uncertainty about when a fixation or a saccade starts and stops. This increased certainty will also translate into less noisy estimates of eye movement measures, such as fixation durations, time to first fixation, and so on.
A study by Andersson, Nyström & Holmqvist (2010) distinguishes between latency measures, which depend on one given and one sampled point, and duration measures, which depend on two sampled points. A latency measure could be, for example, the time to first fixation on a target. This duration or interval must have a start point and an end point. The start point is given by a computer or eye-tracker clock which has a sufficiently high sampling frequency for its sampling error to be disregarded. The end point is determined by the start of the fixation, which is determined by a fixation filter and its calculations. Where the “true” start of a fixation happens will be somewhere inside a sampling interval, but it could be just before the gaze position is sampled (its position is picked up immediately), or just after it has been sampled (and we must wait for the next sample to be registered). This sampling error can be near zero if the eye is sampled just as the fixation starts, but as large as one sampling interval if it happens just after a sample. On average, however, it will be half of one sampling interval, so with a 60 Hz eye tracker, this will be (1000/60)/2 = 8.33 ms, with a minimum and maximum error of 0 and 16.67 ms, respectively. So even if a participant produces several time-to-first-fixation events with identical latencies, their registered latency by the 60 Hz eye tracker will vary by up to 16.67 ms, and add noise of that range to your analysis. This also means that your latency measures will be slightly over-estimated, by the duration of half a sampling interval on average.
For duration measures, such as fixation durations, you have errors at both ends of the event, because both parts are sampled by the eye tracker. The good news is that if you over-estimate the beginning of the fixation, and over-estimate the end of the fixation, both errors could cancel each other out so you get the true fixation duration. This is what happens, on average, so there is no bias in your measurements. However, you now have an error in the beginning and in the end of your measurements, so the duration can both be underestimated by one sampling interval as well as overestimated by one sampling interval. This mean that the sampling error is centered on zero, but the range of the noise goes from -16.67 ms to +16.67 ms for a 60 Hz eye tracker.
In practice, this sampling error will translate into noise and uncertainty in the analysis, making it more difficult to get results that pass some statistical significance or credibility level. What can you do about this? The easy solution is just to get an eye-tracker with sufficiently high sampling frequency for studying your particular phenomenon. If what you are studying is a very brief and noisy phenomenon, then having a high sampling frequency is important, but if the phenomenon shows up as large differences between experimental conditions with low levels of other types of noise, then even a slow eye tracker will work well. It’s a good rule of thumb that the sampling interval of an eye tracker should half be of the duration of the event in question. If you want to cut down on the uncertainty in your data while still not using a faster eye tracker, then it’s possible to collect additional data to compensate for this uncertainty. The rule of thumb (according to Andersson, Nyström & Holmqvist, 2010) is to quadruple the amount of observations you collect to be equivalent to doubling your sampling frequency.
In gaze-contingent research, where what you show on screen is determined by your current gaze position, there is not really any upper limit to how fast you want to sample the eye. There are, however, diminishing returns on the increases in the sampling frequency. You want to be able to sample the eye at high speeds, to know immediately if the eye crosses some boundary on the screen, or produces a saccade towards some target. You need to know this early, because you want to change the information on the screen, even during the saccade so the participant is less likely to notice the change. This means that the faster you sample the eye, the more time you have to identify a saccade and to change the stimulus on the screen. For gaze contingency, other timing issues become important, such as the timing of the stimulus onset on the screen, and the precision with which the eye tracker streams the gaze data.
The pupil produces slow responses, which means that the sampling frequency does not need to be high. For example, the rhythmic hippus movement at 1-2 Hz is detectable by any device filming the eye. Addtionally, pupil size fluctuations driven by changes in mental workload also produce slow responses on the order of seconds (Klingner, Kumar, Hanrahan, 2008).
When synchronizing with other sensors, most notably sensors for electrodermal activity (EDA; called galvanic skin response – GSR), electromyography (EMG), electrocardiography (ECG), and respiration, you may have sensors that produce data at different sampling rates. Depending on your skills in signal processing, there are standard approaches for sample-rate conversions if the goal is to align the data to a common sampling rate (see Wikipedia: Sample-rate conversion).
In Tobii Pro Lab, it is possible to add some sensors directly, like the Shimmer GSR device. Then Pro Lab will take care of the different sample rates and aligning them on a common time line.
Andersson, R., Nyström, M., & Holmqvist, K. (2010). Sampling frequency and eye-tracking measures: how speed affects durations, latencies, and more. Journal of Eye Movement Research, 3(3). doi:http://dx.doi.org/10.16910/jemr.3.3.6
Klingner, J., Kumar, R., & Hanrahan, P. (2008, March). Measuring the task-evoked pupillary response with a remote eye tracker. In Proceedings of the 2008 symposium on Eye tracking research & applications (pp. 69-72). ACM.
Juhola, M., Jäntti, V., & Pyykkö, I. (1985). Effect of sampling frequencies on computation of the maximum velocity of saccadic eye movements. Biological Cybernetics, 53(2), 67-72.
Sample-rate conversion (n.d.). In Wikipedia. Retrieved February 21, 2018, from https://en.wikipedia.org/wiki/Sample-rate_conversion