Tablet Bad Behaviour
Here at Cursor Insight we run motion analytics on a wide range of devices. During the development of our first product, which identifies users through biometric dynamic signatures, we’ve been through heaven and hell to achieve the desired results. One of the bottlenecks was the unreliability of the data received from the devices used to record signatures. In this post I will explore some of the undocumented properties of commercially well-known tablets and signing pads, following the footsteps of Erika Griechisch, Jean Renard Ward and Gergely Hanczár. Their conference paper is available on ResearchGate.
When developing Machine Learning algorithms, the first and foremost important thing is to gather the cleanest data possible. Most tablets and touchscreens are not intended for capturing biometric handwriting data, instead for user interface devices. Most tablets and touchscreens do not provide good data for handwriting biometrics. So we focus on some of the special devices specifically for electronic signing. These are advertised as producing better-quality datasets for handwriting. The devices, however, contain deliberate interventions (smoothing, angle snapping, etc.) in their software to hide the deficiencies of their hardware, to make the data look better. These features (or bugs?) are rarely disclosed, but they can make important performance even worse. The devices’ official performance specs are sometimes far from reality (often important specifications — such as accuracy, resolution — are not given at all). The sampling rates, for instance, were only stated in 2 devices out of the 8 that we thoroughly examined. The accuracy of the measured X-Y coordinates is, in most cases, not included in the devices’ specifications. Our goal was to compare the official specs with our actual test results. The funny part is, in the majority of the cases the vendors’ specifications were so off, that there was not much to compare.
How we carried out the experiment
First of all, this is just one experiment, and shows just a few of the bad behaviors of tablets. Eight tablets, marketed as high-quality writing tablets were tested: most of them are widely used among researchers. The same experiment was conducted on all devices: a roughly one minute long monotonous circular motion data was captured using a strobe-controlled turntable with constant rotational speed. The main benefit of this setup is that the reported timestamps can be verified by calculating the distances between all recorded sample points. If the distances were inaccurate or inconsistent, then either:
a) the position of the coordinates measured are inaccurate, or
b) the measured timestamps are inaccurate.
From our investigations we are concerned that any research study based on raw acceleration data could be questionable. Acceleration was just an example, other time-dependent properties and features should be treated with caution as well. Therefore, understanding the actual behaviours of a particular device and having adequate remediation is essential.
Being able to tell when a given handwriting event occurred (especially in the matter of electronic signatures) is a must. For this reason, some vendors include timestamps with each data point. However, many devices report unreliable timestamps. When there is a difference between the clock time and the actual sampling time, the problem may be that the vendor rounded the timestamp number, or the timestamps are affected by irregular buffering delays, are measured at the wrong point in the firmware algorithms, or the timestamps are simply buggy. Luckily, most devices measure position at constant intervals, and we can calculate the time from the sampling rate.
Speed & Reporting
Based on the turntable-setup, the measurement of constant velocity was expected. After seeing the results however, it became clear that there is much noise to the signals. Raw measurements alone were unreliable and further altered by the devices’ drivers, resulting in changes in measurement. For instance, in one of the devices, because of a minor calibration fault, timestamp differences are not constant. I.e. the time reported between two data points is not equal. Therefore, when measuring handwriting, speed cannot be well established based on the device timestamps. Another device, however, has 4 different interval values instead of two. See the figures below.
Take tablet data with a large grain of salt
Do not take any measurement for granted for signature analysis. The commercially available (see more information in the publication below) tablets and signing pads can still form a reliable basis for movement analysis and research. However, one should always double-check and fix the data before blindly applying algorithms on them to extract information.
The article has been posted and is available for read here.
The original publication can be found on ResearchGate as well. Erika Griechisch was also representing Cursor Insight in Darmstadt. As references, see E. Griechisch, J. R. Ward and G. Hanczár, “Anomalies in measuring speed and other dynamic properties with touchscreens and tablets,” 2019 International Conference of the Biometrics Special Interest Group (BIOSIG), Darmstadt, Germany, 2019, pp. 1–6.