Quantitative comparison of a mobile, tablet-based and two stationary, video-based eye-trackers

The analysis of eye-movements is a non-invasive, reliable and fast method to detect and quantify brain (dys-)functions. Here, we investigated the performance of two novel eye-trackers: the Thomas-Oculus-Motus-research mobile (TOM-rm) and the TOM-research stationary (TOM-rs) and compared it with the performance of a well-established video-based eye-tracker, i.e., the EyeLink 1000 (EL). The TOM-rm is a fully integrated, tablet-based mobile device that presents visual stimuli and records head-unrestrained eye-movements at 30Hz without additional infrared (IR) illumination. The TOM-rs is a stationary, video-based eye-tracker that records eye-movements at either high spatial or high temporal resolution. We compared the performance of all three eye-trackers in two different behavioral tasks: pro– and anti-saccade and free viewing. We collected data from human subjects while running all three eye-tracking devices in parallel. Parameters requiring a high spatial or temporal resolution (e.g., saccade latency or gain), as derived from the data, differed significantly between the EL and the TOM-rm in the pro– and anti-saccade task. In the free viewing task, larger noise and the lower frame rate of the TOM-rm caused deviations of the results with respect to the EL. Differences between results derived from the TOM-rs and the EL were most likely due to experimental conditions, which could not be optimized for both systems simultaneously. We conclude that the TOM-rm can be used for measuring eye-movements reliably at comparably low spatial and temporal resolution. The TOM-rs, on the other hand, can provide high-resolution oculomotor data at least on a par with an established reference system.


Introduction
accordance with the Declaration of Helsinki and were approved by the local ethics 96 committee (AZ-2012-23K). 97

Eye-tracker and laboratory setup 98
The laboratory setup for concurrent data collection with all three eye-trackers is shown 99 in Figure 1. (e.g., Full-HD @ 150Hz) or high temporal (e.g., 640 x 480px @ 2000Hz) resolution. Here, 114 we measured with a frequency of 500Hz and a resolution of 640 x 480px. This relatively 115 low frame rate was necessary to compensate for the lower IR light sensitivity required by 116 the EL. The TOM-rs camera was placed to the left and below the participant's head next 117 to the centrally positioned EL. Both, the TOM-rs and the EL, need an individually adjusted 118 IR illumination. The EL requires a less intense IR light-source, because it uses a lens with a 119 fixed focal length but high light intensity. The TOM-rs on the other hand uses a zoom lens 120 with a variable focal length between 16mm-300mm, and hence requires IR-illumination 121 at a higher intensity. To be able to measure with all eye-trackers simultaneously, we had 122 to adjust both IR illuminators (EL and TOM-rs) such that both eye-tracking systems could 123 reliably detect the participants' pupil. The EL-system was set to an intensity of 75% of its 124 infrared illumination to avoid over-illumination by the additional IR light source of the 125 TOM-rs. The infrared intensity of the TOM-rs illuminator was adjusted by modifying the 126 angle of the irradiated infrared light in a subject-specific manner. This necessary tradeoff, 127 however, has the potential to compromise overall data-quality due to non-optimal 128 lighting conditions (see also Discussion). 129 order (Figure 2a). At the beginning of each trial, a blue or red fixation point (FP) with a 140 diameter of 1° visual angle was displayed on a grey background in the center of the screen. 141 The participants were asked to fixate this target. The color of the fixation target (blue or 142 red) indicated whether a pro-or anti-saccade should be performed during the further 143 course of the trial (blue: pro-; red: anti-saccade). After 1s the FP was switched off. 200ms 144 later, a white target point (TP) with a diameter of 1° visual angle appeared for one second 145 in pseudo-randomized order 10.8° to the right or left of the FP on the horizontal meridian 146 (HM). The subjects were instructed to perform the pro-or anti-saccade as quickly as 147 possible. Between trials, a grey screen was presented for 1 second. In total, this paradigm 148 took about five minutes. 149

Free Viewing Task 150
In the second task, static images were displayed on the screen for a duration of 6s ( Figure  151 2b). 152  (Table 1). 226 In the ideal case, the yield should be 1.0, since each trial should trigger two saccades, i.e., 235 the saccade towards the saccade goal and a saccade back to the central fixation. Yet, since 236 there were trials in which correction saccades occurred, or in which subjects did not 237 return their gaze back to the central fixation with a single saccade, or in which noise was 238 detected as a saccade, often more than two saccades were detected per trial, possibly 239 resulting in a yield > 1. 240 In the free viewing task only saccades larger than 1° amplitude were included for all three 241 eye-tracker, because saccades smaller than 1° could not always be reliable detected with 242 the TOM-rm. 243

Fixations 244
For the individual evaluation of the data, the detection of the fixations for the EL was 245 performed with the built-in detector from the manufacturer. Here, all eye traces that do 246 not qualify as a saccade or smooth pursuit are considered a fixation. The detection of 247 fixations for the TOM-eye-trackers was performed using the EyeMMV toolbox. The lower 248 limit of the fixation time for both TOM-eye-trackers was set to 150ms. In addition, in this 249 toolbox, two spatial thresholds t1 and t2 have to be defined, which determine the 250 maximum eye position jitter allowed to consider an eye trace as fixation (Krassanakis et  251 al., 2014). We used values of t1 = 1° and t2 = 0.7° for evaluation with the TOM-rm data 252 and t1 = 0.5° and t2 = 0.3° for evaluation of the TOM-rs data. For the same evaluation 253 applied to data from all three eye-trackers, we used the fixation detector of the EyeMMV 254 toolbox with the thresholds t1 and t2 of the TOM-rm. Before, the eye-position data of the 12 pro-and anti-saccade task of the TOM-rm were filtered using a median filter over 0.3s, 256 since they are subject to high noise. The TOM-rs eye-position data were also filtered over 257 0.2s due to the noise resulting from the suboptimal illumination condition. 258 We also filtered the eye-position of the TOM-eye-tracker of the free viewing task with a 259 median filter (TOM-rm & TOM-rs same evaluation: 0.2s; TOM-rs individual evaluation: 260 0.05s). 261

Pupillary response 262
The pupillary response was only recorded with the EL and the TOM-rs systems. The light 263 and dark reflex of the pupil were determined in the free viewing task during the 264 observation of the light and dark images. The light reflex is typically considered the 265 beginning of the constriction of the pupil during viewing bright images. In contrast, the 266 dark reflex is considered the beginning of the pupil dilation while viewing dark images. In We did not calculate these two parameters for the dilation, because due to the slow dark 283 adaptation the average pupil area was about to reach a maximum value only at the end 284 of a given trial (Wang & Munoz, 2014). We also aimed to quantitatively compare the 285 dynamics of the pupil area as measured from the stationary eye-trackers. Therefore, we 286 averaged the size of the pupil area over all subjects and determined the 95% confidence 287 interval, separately for bright and dark images. 288

Blink detection 289
Blink's artifacts and all data points 60ms around a blink were excluded from further 290 analysis. For the individual evaluation of the EL data we employed the built-in eyelid blink 291 detector, where blinks are defined as periods of data for which the pupil cannot be 292 detected. For the TOM-rs we first applied a median filter (width: 115samples) on the pupil 293 area data. All samples of this new time series data that deviated by more than a specific 294 value, which we determined for every participant's dataset individually, were considered 295 blinks. In the same way, we detected the blinks of the EL for the same evaluation. Due to 296 the fact that the TOM-rm does not make use of IR-illumination, no pupillometry data were 297 available for the TOM-rm, but the Eye-Aspect-Ratio (EAR) (Cech & Soukupova, 2016). With 298 the EAR, blinks were detected in the same way as with the TOM-rs, but with a larger width 299 of the samples of the moving average (width: 35 samples, corresponding to > 1s). This 14 long interval was necessary because of the relatively high noise in the EAR data. 301 Accordingly, only a rough measure of the EAR was created. Both the blinks in the pupil 302 area data and in the EAR were expressed by short, extreme peaks in the time series data. 303 Finally, saccades were only included in our data analysis if no blink was detected from 304 200ms before until 200ms after the saccade. 305

Statistics 306
For the statistical analysis, we performed a paired-sample t-test to probe for differences 307 in the quantitative eye movement parameters derived from data recorded with the three 308 eye-trackers. Differences were considered significant if the p-value was smaller than α = 309 0.05. We also performed a Bonferroni correction, because for a given dataset, we 310 performed more than one statistical test. This is why p-values had to be corrected for 311 multiple comparisons by multiplying the uncorrected values with the number of 312 hypotheses (pro-/anti-saccade task: 6 hypotheses, free viewing task: 3 hypotheses), 313 resulting in new p-values. Consequently, this approach allows us to keep the typical alpha 314 level, indicating significant differences, at 0.05. 315 In this study we used a rather unconventional method to determine the standard 316 deviation (SD) in order to probe for differences in eye position data between the eye-317 trackers. We used this method detailed below because the inter-individual variance, i.e., 318 variance of data across subjects, could potentially mask the differences in the data caused 319 by the eye-trackers. We determined the SD as follows: First, for each subject, the mean 320 value across all three eye-trackers of a given parameter ̅ was determined. I.e., if the 321 saccade latency for one subject is for the EL, for the TOM-rm and for 322 the TOM-rs, then we determined the mean value ̅ as follows: The mean value per subject per eye-tracker , − and − was subtracted 325 from the mean value over all eye-trackers ̅ 326 Then the SD was calculated over these values. In the following we called this the inter-330 is the i-th value in the data set, ̅ is the mean value of the data set and n is the sample 333 size. 334

335
Here, we compared the performance of two novel eye-trackers, the TOM-rm and the 336 TOM-rs, with a well-established eye-tracker in the field, the EyeLink 1000 (EL). Typical eye 337 movement parameters were analyzed in two different behavioral tasks (pro-and anti-338 saccade and free viewing). In a first step, we analyzed data from each eye-tracker 339 individually, i.e., with software either provided by the manufacturer (EL) or customized to 340 the specifics of the eye-tracker (TOM-rm and TOM-rs), called individual evaluation in the 341 following. In a second step, we analyzed data from all three eye-trackers with the same 342 data analysis programs, i.e., those developed for the TOM-rm. We refer to this approach 343 as same evaluation in the following. All mean values, inter-device SD and inter-individual 344 SD of the pro-and anti-saccade task can be obtained from the supplements in Table S1. 345 The respective results of the free viewing task are provided in Table S2.

Pro-and anti-saccade task 347
The error rate of pro-saccades (trm(20) = -1.40, trs(20) = -1.40, prm = 1.00, prs = 1.00) and 348 anti-saccades (trm(20) = -1.38, trs(20) = -1.92, prm = 1.00, prs = 0.42) as determined with the 349 individual evaluation were not significantly different between the data obtained with the 350 EL and the TOM-rm (as indicated by prm values) or the EL and the TOM-rs (as indicated by 351 the prs values) (Figures 4a,b). Only correct pro-and anti-saccades were used for computing 352 the saccadic gain. We observed a significant difference between data from the EL and 353 TOM-rm concerning the gain in the pro-and anti-saccade task (pro: trm(20) = 6.20, prm = 354 2.79 x 10 -5 , anti: trm(20) = 7.40, prm = 2.30 x 10 -6 ). On average, the gain as determined from 355 the TOM-rm data was 7% smaller as compared to the gain derived from the EL-dataset. 356 There were no such differences between data obtained with the EL and the TOM-rs (pro: 357 trs(20) = 1.30, prs = 1.00, anti: trs(20) = 1.56, prs = 0.81. Figure 4c). For the pro-saccade 358 latency, we found a significant difference between the EL and the TOM-rm datasets, but 359 not between the EL and the TOM-rs data-sets (trm(20) = 13.20, trs(20) = -1.43, prm = 1.49 x 360 10 -10 , prs = 1.00). This was also the case for the anti-saccade latency (trm(20) = 10.75, trs(20) 361 = -2.59, prm = 5.53 x 10 -9 , prs = 0.10. Figure 4d). In both cases, pro-and anti-saccades, 362 saccadic latencies derived from the TOM-rm dataset were about 18ms shorter than those 363 values determined from the EL-dataset. Figure 4e shows the average number of saccades. 364 We found no significant differences between results from the eye-trackers concerning the In a second step, we compared the performance of the eye-trackers by analyzing the data 384 with the same evaluation approach. For this purpose, we used the routines as developed 385 for analyzing the TOM-rm data, since this eye-tracker provides data with the lowest 386 spatial and temporal resolution. To make results comparable, we first down-sampled EL 387 and TOM-rs data to 30Hz, i.e., the genuine sample rate of the TOM-rm (see Methods for 388 details). No difference was found for the pro-saccade error rates neither between the EL 389 and TOM-rm datasets nor between the EL and TOM-rs datasets (Figure 5a). Likewise, we did not find significant differences for the error rates in the anti-saccade task for the TOM-TOM-rs were the same. There was a significant difference of about 3%, though, between 393 the EL and the TOM-eye-trackers in the pro-and anti-saccade gain (pro: trm(20) = 9.29, 394 trs(20) = 3.97, prm = 6.43 x 10 -8 , prs = 4.55 x 10 -3 ; anti: trm(20) = 6.89, trs(20) = 4.00, prm = 6.46 395 x 10 -6 , prs = 4.20 x 10 -3 ; Figure 5c). Also here, the gain as determined from the TOM-rm 396 data was about 7% smaller than its corresponding value determined from the EL-dataset. we found no significant difference between the two stationary eye-tracker datasets The spatial accuracy of the TOM-rm was slightly worse than one degree of visual angle, 425 therefore only saccade amplitudes greater than one degree of visual angle were 426 considered in the evaluation of the free-viewing data. 427 To analyze the eye-movements of the free viewing task we first determined the mean 428 number of fixations per image (stimulus duration: 6s) with the individual evaluation 429 approach. We found a significant difference between the EL dataset and the dataset from 430 the TOM-rm concerning the number of fixations (roughly 12 fixations derived from the number of fixations of the TOM-rs was also smaller than that of the EL (trm(20) = 6.63, 434 trs(20) = 2.65, prm = 5.54 x 10 -6 , prs = 4.62 x 10 -2 ; Figure 6a). Furthermore, we found a 435 significant difference between the EL and the TOM-rm datasets concerning the mean 436 amplitude of saccades, but not between the datasets of the two stationary eye-trackers. 437 Saccade amplitudes as determined from the TOM-rm datasets were on average approx. 438 15% smaller than those obtained from the EL (trm(20) = 11.06, trs(20) = 1.00, prm = 1.69 x 439 10 -9 , prs = 0.99. Figure 6b). In the mean fixation duration, there was no significant 440 difference between the EL and the TOM-eye-tracker datasets (trm(20) = 0.69, trs(20) = 0.81, 441 prm = 1.00, prs = 1.00; Figure 6c). 442 In a second step, we examined the eye movement parameters in the free viewing task 443 using the same evaluation routines to identify potential differences between the three  image. The pupil area was normalized by forming the z-score and this value was set to 0

522
In this study we quantitatively compared eye-movement parameters as derived from 523 measurements with two novel eye-trackers (TOM-rm and TOM-rs) with those from the 524 EyeLink 1000 (EL). One of these new eye-trackers is a high-resolution stationary eye-525 tracker (TOM-rs) and the other is a mobile, tablet-based eye-tracker with a frame rate of 526 30Hz (TOM-rm). The EL is a well-established eye-tracker in oculomotor research and 527 therefore served as a reference system for our study. 528 trackers, which has advantages and disadvantages. One of the advantages was that we 531 were able to compare the identical eye movement parameters quantitatively and thus 532 directly compared the eye-trackers with each other. The disadvantage of this method was 533 that the different eye-trackers have different demands especially concerning the lighting 534 of their environment, so compromises must be made. Recordings with the EL and TOM-535 rs require IR-light and should take place in a darkened room so that the contrast between 536 pupil and iris increases. Since the TOM-rm requires visible light, our measurement took 537 place in a well-lit room, which can slightly affect the IR spectrum of the EL and TOM-rs 538

Systems (Kunka & Kostek, 2009). 539
Ideally, the results as obtained from the two stationary eye-trackers (EL and TOM-rs) 540 should not differ significantly neither concerning eye movement data nor results from 541 pupillometry. However, both systems require IR illumination and their IR light sensitivities 542 are different. Our highest priority was to concurrently measure eye-movements with all 543 three systems. Only such an approach allows to compare identical oculomotor data with 544 each other. As a consequence, we had to find a compromise concerning the IR 545 illumination. Due to the fixed focal length lens but high light sensitivity, the required IR 546 illumination for the EL is weaker than for the TOM-rs, which uses a variable focal length 547 zoom lens with slightly lower light sensitivity. Our intention was to keep the measurement 548 conditions between the eye-trackers as similar as possible. Thus, we decided to use an EL-549 optimal IR illumination, because otherwise the images from the EL would have been 550 overexposed and non-usable. On the other hand, the TOM-rs allows to flexibly adjust the 551 sampling rate and the shutter speed of the camera, which allowed us to find the best 552 setting of this non-optimal illumination. Yet, this compromise still affected the eye-553 tracking quality of the TOM-rs, since these images tended to be underexposed and the 554 resulting noise could only partially be eliminated by smoothing the data. 555 For the quantitative comparison, we first evaluated the data of the three eye-trackers 556 with individually adapted evaluation software. In this type of evaluation, both hardware-557 specific and software-specific components were implicitly included in the results. To 558 understand which differences are due to the hardware, we used the same evaluation 559 software (that of the TOM-rm) for all three eye-trackers. 560

Functional characteristics of the two novel eye-trackers 561
The TOM-rm is a fully integrated mobile device, which means that a laboratory 562 environment is not necessary for a measurement. Another characteristic of the TOM-rm 563 is that the device is lightweight, easy to use and requires only a short training in handling. 564 Unlike the EL or TOM-rs, the quality of gaze detection depends especially on the lighting 565 condition (in the visible light range). The TOM-rm measures eye-movements at a frame 566 rate of 30Hz, which means that parameters that need a high temporal resolution (e.g., 567 saccade latency) cannot be reliably determined. Figure 10 shows, from a theoretical 568 perspective, the effect of high noise and low sampling frequency in comparison to low 569 noise and high sampling frequency on the fixation and saccade detection. Figure 10  Since the sharpest contour is between the iris and pupil and not between the iris and 581 sclera, the pupil can be detected much more precisely with IR light than the iris with visible 582 light (Kunka & Kostek, 2009). In the raw data the noise of the TOM-rm was relatively high 583 as compared to the stationary eye-trackers. Accordingly, it appears challenging to 584 determine oculomotor parameters, which require a high spatial resolution (e.g., saccade 585 gain). Our results confirm this view and suggest that the TOM-rm is suitable for the 586 measurement of oculomotor parameters that don't require high spatial or temporal 587 resolution, like the pro-and anti-saccade error rate, and measurements that require a 588 flexible setting for the recordings. More accurate measurements, on the other hand, are 589 possible with the TOM-rs. With the high-resolution camera, which can record data with 590 up to 2000Hz, it is possible to perform temporally and spatially accurate and precise 591 measurements. Since the TOM-rs uses a zoom lens with a variable focal length between 592 the error rate of the EL was smaller (not significant) than that of the TOM-rs, which was 616 due to the fact that the previously mentioned fast erroneous saccades were not detected by the EL saccade detector in a few cases. Typically, saccades undershoot the target 618 position (Hallett, 1978;Smit, Van Gisbergen, & Cools, 1987;Krappmann, 1998). Our data 619 are in line with these previous results. 620 Saccadic gain as derived from the TOM-rm datasets, however, were on average approx. 621 7% smaller than the respective values derived from the EL datasets. We assume that this 622 difference was related to the lower spatial and temporal resolution, as shown exemplary 623 in Figure 11 shown for the EL (purple), the TOM-rm (orange) and the TOM-rs (yellow) during a saccade. 625 The blue symbols represent the start of a saccade and the red the end. The square 626 corresponds to the saccade start/end of the EL, the circles to the TOM-rm and the crosses 627 to the TOM-rs. The saccade of the TOM-rm is smaller than that of the stationary eye-628 tracker, thus the saccade gain is also smaller. 629 When evaluating data from all three eye-trackers with the same evaluation, saccadic gain 630 as derived from the TOM-rs data was significantly smaller than the value derived from the 631 EL data set. We assume that this was due to the noisier TOM-rs data under non-optimal 632 IR illumination. The gain in the TOM-rs dataset turned out to be reduced only for the same 633 evaluation. It appears likely that this seemingly different performance was induced by the 634 down sampling of the dataset. This had a bigger effect on the TOM-rs data due to the 635 compromise in the intensity of the IR illumination than on the EL data. Figure 11(c/d) 636 shows the horizontal eye position over time for the EL (c) and for the TOM-rs (d). The data 637 set with a frame rate of 500Hz is represented by the solid line and the down sampled data 638 (30Hz) by the dashed line. The blue symbols represent the start of a saccade and the red 639 the end. The cross corresponds to the saccade start/end of the EL/TOM-rs at 500Hz and 640 the circles at 30Hz. The compromise reduces the precision, which in turn could reduce the 30 shows that the gain as derived from the two stationary eye-trackers was almost the same, 643 but not for the mobile eye-tracker. 644 For the analysis with the same evaluation algorithm, the data first was down-sampled to 645 30Hz. Anti-saccade latency for all three eye-trackers was higher than for pro-saccades, 646 which corresponds to findings reported in the literature (Everling & Fischer, 1998). The 647 low frame rate and high noise of the TOM-rm results in lower saccade latency than 648 stationary eye-trackers, this effect is shown in Figure 10. Similar to the saccade gain, down 649 sampling to 30Hz affected the latency of the TOM-rs-data more than the EL-data. Figures  650 11(c)/(d) show that for the down sampled data the saccade start of the TOM-rs was 651 detected earlier than for the non-down sampled data and the EL data. 652 The number of saccades as derived from the TOM-rm dataset was on average 5% smaller 661 than that from the EL dataset. This difference was mainly a consequence of the saccade 662 duration threshold (saccade duration < 200ms). In Dalmaijer (2014), it was shown that the 663 saccade duration derived from a low-frequency eye-tracker was higher than that from a 664 high-frequency eye-tracker. Nevertheless, we wanted to restrict saccade duration to this 665 upper limit because even larger values appeared non-physiological. On average, (11.00  666 7.98) saccades per subject were removed for the TOM-rm dataset using this saccade 667 duration criterion. Concerning the same evaluation (6.05  5.29) saccades per subject of 668 the EL datasets were removed and (5.24  6.60) saccades per subject of the TOM-rs 669 datasets. 670 The duration of the fixation in the same evaluation was shorter for the TOM-rm than for 671 the two stationary eye-trackers, because small saccades of the stationary eye-trackers in 672 the individual evaluation as shown in Figure 12(a) did not interrupt periods of fixations as 673 determined in the same evaluation (Figure 12(b)). 674 were no small saccades during fixation, then the saccades of the stationary eye-trackers 685 were longer than those of the TOM-rm, due to the higher frame rate. However, if small 686 saccades interrupt the fixations of the stationary eye-trackers, then these fixations were 687 shorter than those of the TOM-rm. Considering the total number of trials of the subjects, 688 the differences compensate each other and lead to the result that on average there was 689 no difference between the stationary and the mobile eye-trackers. 690 In a last step we determined all parameters of the pro-and anti-saccade task with the 691 TOM-rs under ideal IR-illumination conditions in seven subjects. As shown in Figure 8, 692 noise was greatly reduced in this optimized condition in contrast to the measurement 693 under compromise conditions. We had assumed that under ideal IR-illumination 694 conditions the saccade gain would be closer to the values previously derived from the EL-695 dataset. Yet, unexpectedly, this was not the case, as shown in Figure 9(c). Given that these 696 data were recorded from a new cohort of subjects, we speculate that these subjects on 697 average had a lower saccadic accuracy. 698

Free viewing task 699
The investigation of eye-movements during the exploration of images can be a powerful 700 tool for the detection of neurodegenerative diseases. In Parkinson's disease, for example 701 the saccade amplitudes are on average smaller than in healthy age-matched controls 702 (Matsumoto et al., 2011;Matsumoto et al., 2012). In this study we investigated the 703 performance of the TOM-eye-trackers during a free viewing task with 30 different images two stationary eye-tackers (Figure 13(a)). This effect also occurred in the pro-and anti-709 saccade task. 710 When we determined the fixation duration with the same evaluation, the duration was 711 lower for the EL and TOM-rs than for the individual evaluation. This was mainly since the 712 spatial range (t1 and t2) for the fixation detector of the TOM-rm was chosen relatively 713 high (see Methods for details). This range was obviously too high for the TOM-rs and EL 714 datasets and resulted in fixations that were interrupted by small saccades, which however 715 were not detected separately. Accordingly, the individual evaluation was better suited for 716 the two stationary eye-trackers. If we compare the TOM-rs with the EL for the evaluation 717 with the individual and same fixation detector, we found no difference between the eye-718 trackers. Values derived from the TOM-rm did not differ from those from the two 719 stationary eye-trackers concerning fixation duration in the individual evaluation, but for 720 the same evaluation. This also suggests that the fixation detector in the same evaluation 721 was not suitable for the stationary devices. 722 For a given temporal interval, the number of fixations is inversely proportional to the 723 fixation duration. If the number of detected fixations increases, the duration of the 724 detected fixations must decrease. This is implicitly reflected in the data shown in Figure  725 13(c/d). 726 In the same evaluation we found a significant difference between the TOM-rm and the 727 stationary eye-tracker data. Like in the pro-and anti-saccade task, the main reason was 728 that interruptions due to small saccades could not be detected because of the higher 729 noise in the TOM-rm data. Not only eye-movements but also the pupillary response can be an important biomarker 745 for neurological diseases, as shown in Wang et al. (2016). In this study we investigated the 746 difference in pupillary response between the two stationary eye-trackers while viewing 747 bright and dark images. According to the literature, the latency of the constriction onset 748 of the light reflex is in the range of 230 to 357ms and about 445ms for the dark reflex and dark reflex showed similar values. In the light and dark reflex, the EL and the TOM-rs 751 datasets did not differ significantly. The time-course of the pupil areas while viewing the 752 bright images of the two eye-trackers did not show any difference, as was the case for the 753 contraction amplitude for the TOM-rs. However, the constriction amplitude of the TOM-754 rs was lower, however not significant, than of the EL. Like for the eye movement tasks, 755 the differences probably were caused by the compromise in IR illumination. Without 756 optimal illumination the noise increases and pupil detection becomes less reliable. 757

758
We conclude that simultaneous measurement with three eye-trackers with different 759 demands on the environment, requires compromises that affect the data quality of the 760 eye-trackers. Nevertheless, this compromise is worth it, because it allows a direct and 761 quantitative comparison between the eye-trackers. Our results show that the mobile 762 tablet-based eye-tracker (TOM-rm) is suitable for certain basic oculomotor experiments 763 like the pro-and anti-saccade task. Experiments in which parameters are collected that 764 require a high frame rate or a high spatial accuracy or precision, such as saccade-latency 765 or gain require eye tracking systems with a high spatial and temporal resolution. Our 766 results show that -given ideal IR illumination -the quality of oculomotor data obtained 767 from the TOM-rs are on a par with those obtained from a reference in the field, i.e., the 768 EyeLink 1000. 769 experiment after instruction and training by A.K.. A.K. performed the analyses of the data. 775 The manuscript was written by A.K., S.D. and F.B.. The two TOM eye-trackers are 776 developed and distributed by Thomas RECORDING GmbH (Giessen, Germany). S.D. was 777 involved in the development of the two TOM eye-trackers as the chief scientific officer of 778 Thomas RECORDING and has been paid by Thomas RECORDING during that time. As part 779 of the collaborative research project DIADEM the eye-trackers were freely provided to the 780 Dept. Neurophysics (Philipps-Universität Marburg, Germany). 781

782
The raw data supporting the conclusions of this article will be made available by the 783