US 12,229,992 B2
Method of performing a user-specific and device-specific calibration of point of gaze estimation
Karol Duzinkiewicz, Banino (PL); Jan Glinko, Gdansk (PL); Artur Skrzynecki, Wejherowo (PL); and Michael Schiessl, Berlin (DE)
Assigned to eye square GmbH, Berlin (DE)
Filed by Eye Square GmbH, Berlin (DE)
Filed on Jun. 30, 2023, as Appl. No. 18/345,911.
Prior Publication US 2025/0005790 A1, Jan. 2, 2025
Int. Cl. G06T 7/73 (2017.01); G06F 3/01 (2006.01)
CPC G06T 7/74 (2017.01) [G06F 3/013 (2013.01); G06T 2207/20081 (2013.01); G06T 2207/20084 (2013.01); G06T 2207/30196 (2013.01)] 12 Claims
OG exemplary drawing
 
1. A method of calibrating point of gaze estimation, comprising the steps:
providing a calibration application;
providing a user device;
wherein said user device comprises a display and a user-facing camera;
wherein said calibration application runs on said user device;
wherein said calibration application causes to be displayed on said display a series of calibration markers;
prompting a user to look at said series of calibration markers by displaying said series of calibration markers on said display;
capturing one or more images of said user as said user is prompted to look at said series of calibration markers;
wherein a series of coordinates on said display is recorded for each of said series of calibration markers;
wherein said series of coordinates comprise a series of target ground truth positions for said series of calibration markers, which comprise a target ground truth position data;
generating a calibration data set by matching said target ground truth position data with said captured images;
processing said calibration data set by a Point of Gaze (“PoG”) estimation pipeline to generate PoG output data; wherein said target ground truth position data and PoG output data comprise a processed calibration data set;
splitting, by a calibration data set splitter, said processed calibration data set into a training calibration data set and a validation calibration data set;
processing, by a support vector regression calculator, said training calibration data set by utilizing one or more SVR training algorithms;
wherein said one or more SVR training algorithms are used to generate one or more SVR models after processing said training calibration data set;
calculating a preliminary positional gaze data for each of said one or more SVR models by utilizing said validation calibration data set;
comparing said preliminary positional gaze data against said validation calibration data set to calculate a mean absolute error for each data point of said validation calibration data set; and
wherein a best SVR model is selected from said one or more SVR models based on which of said SVR model has a lowest mean absolute error.