Training of a gaze tracking model

1. A gaze tracking method comprising:

obtaining training data, wherein the training data comprises:

a first image of an eye captured by an imaging device;

reference gaze data indicating a gaze point towards which the eye was gazing when the first image was captured; and

images of an eye captured at a point in time by first and second imaging devices, respectively; and

training a gaze tracking model using the training data,

wherein the gaze tracking model is adapted to predict a gaze ray of an eye using an image of that eye captured by an imaging device,

wherein the training comprises:

forming a distance between said gaze point and a gaze ray predicted by the gaze tracking model using said first image;

forming a consistency measure based on at least a gaze ray predicted by the gaze tracking model using the image captured at said point in time by said first imaging device and a gaze ray predicted by the gaze tracking model using the image captured at said point in time by said second imaging device;

forming an objective function based on at least the formed distance and the consistency measure; and

training the gaze tracking model using the objective function.

2. The method of claim 1 , wherein the imaging device by which said first image was captured is said first imaging device.

3. The method of claim 1 , wherein said first image is one of the images captured at said point in time.

4. The method of claim 1 , wherein the eye in said first image is the same eye as in the images captured at said point in time.

5. The method of claim 1 , wherein said first image was captured while the eye in said first image was gazing towards a known reference point, and wherein the reference gaze data indicates the known reference point.

6. The method of claim 1 , wherein the consistency measure indicates a degree of consistency between gaze rays predicted by the gaze tracking model using images captured by at least the first and second imaging devices.

7. The method of claim 1 , further comprising:

obtaining a further image of an eye captured by an imaging device; and

predicting a gaze ray using the further image and the trained gaze tracking model.

8. The method of claim 7 , wherein the eye in the further image is the same eye as in the images captured at said point in time.

9. The method of claim 1 , wherein the gaze tracking model predicts a gaze ray in the form of at least:

two points along the gaze ray; or

a point along the gaze ray and a direction of the gaze ray.

10. The method of claim 1 , wherein the gaze tracking model is adapted to estimate a reference point of an eye along the gaze ray, and wherein forming the consistency measure comprises:

forming a distance between the reference point estimated for the image captured by said first imaging device at said point in time and the reference point estimated for the image captured by said second imaging device at said point in time.

11. The method of claim 1 , wherein the gaze tracking model is adapted to estimate a gaze direction of an eye, and wherein forming the consistency measure comprises:

forming a difference between the gaze direction estimated by the model for the image captured by said first imaging device at said point in time and the gaze direction estimated by the model for the image captured by said second imaging device at said point in time.

12. The method of claim 1 , further comprising:

for the eye in the images captured at said point in time, estimating a position of the eye based on at least the images captured at said point in time,

wherein the gaze tracking model is adapted to estimate a reference point of the eye along the gaze ray wherein the training of the gaze tracking model is performed subject to a condition that the estimated reference point should be within a certain distance from the estimated position of the eye.

13. The method of claim 1 , wherein forming the consistency measure comprises:

predicting a first gaze ray using the gaze tracking model and the image captured at said point in time by the first imaging device transforming the first gaze ray to a coordinate system associated with the second imaging device, and forming the consistency measure using the transformed first gaze ray and a gaze ray predicted by the gaze tracking model using the image captured at said point in time by the second imaging device; or

predicting a first gaze ray using the gaze tracking model and the image captured at said point in time by the first imaging device, predicting a second gaze ray using the gaze tracking model and the image captured at said point in time by the second imaging device transforming the first and second gaze rays to a common coordinate system, and forming the consistency measure using the transformed first and second gaze rays.

14. The method of claim 1 , wherein forming the consistency measure comprises:

compensating for a difference between a property of the first imaging device and a property of the second imaging device.

15. The method of claim 1 , wherein the gaze tracking model is a machine learning model.

16. The method of claim 1 , wherein training the gaze tracking model using the objective function comprises:

modifying or extending the gaze tracking model for reducing a value of the objective function; or

modifying or extending the gaze tracking model for increasing a value of the objective function.

17. The method of claim 1 , wherein the image captured at the point in time by the first imaging device is captured from a different angle than the image captured at the point in time by the second imaging device.

18. A gaze tracking system comprising processing circuitry configured to:

obtain training data, wherein the training data comprises:

a first image of an eye captured by an imaging device;

reference gaze data indicating a gaze point towards which the eye was gazing when the first image was captured; and

images of an eye captured at a point in time by first and second imaging devices, respectively; and

train a gaze tracking model using the training data,

wherein the gaze tracking model is adapted to predict a gaze ray of an eye using an image of that eye captured by an imaging device,

wherein the processing circuitry is configured to train the gaze tracking model by at least:

forming a distance between said gaze point and a gaze ray predicted by the gaze tracking model using said first image;

forming a consistency measure based on at least a gaze ray predicted by the gaze tracking model using the image captured at said point in time by said first imaging device and a gaze ray predicted by the gaze tracking model using the image captured at said point in time by said second imaging device;

forming an objective function based on at least the formed distance and the consistency measure; and

training the gaze tracking model using the objective function.

19. The gaze tracking system of claim 18 , wherein

the gaze tracking model is adapted to estimate a reference point of an eye along the gaze ray, the processing circuitry being configured to form the consistency measure by at least:

forming a distance between the reference point estimated for the image captured by said first imaging device at said point in time and the reference point estimated for the image captured by said second imaging device at said point in time; and/or

the gaze tracking model is adapted to estimate a gaze direction of an eye, and wherein the processing circuitry is configured to form the consistency measure by at least:

forming a difference between the gaze direction estimated by the model for the image captured by said first imaging device at said point in time and the gaze direction estimated by the model for the image captured by said second imaging device at said point in time.

20. A non-transitory computer-readable storage medium storing instructions which, when executed by a gaze tracking system, cause the gaze tracking system to:

obtain training data, wherein the training data comprises:

a first image of an eye captured by an imaging device;

reference gaze data indicating a gaze point towards which the eye was gazing when the first image was captured; and

images of an eye captured at a point in time by first and second imaging devices, respectively; and

train a gaze tracking model using the training data,

wherein the gaze tracking model is adapted to predict a gaze ray of an eye using an image of that eye captured by an imaging device,

wherein the training comprises:

forming a distance between said gaze point and a gaze ray predicted by the gaze tracking model using said first image;

forming a consistency measure based on at least a gaze ray predicted by the gaze tracking model using the image captured at said point in time by said first imaging device and a gaze ray predicted by the gaze tracking model using the image captured at said point in time by said second imaging device;

forming an objective function based on at least the formed distance and the consistency measure; and

training the gaze tracking model using the objective function.