Question about iris depth experience

[DavitMob]

Hi,

First of all, your work helped me a lot, thanks for that. In iris depth estimation, the error increases when the person tilts his/her face. Is there something mediapipe deos, or you might have an idea what we can do in this case to decrease the error?

[Rassibassi]

Hi,

glad my code is useful to you.

The depth estimation uses the intercept theorem, where it assumes that the real world iris plane of the user is parallel to the screen/camera plane (in other words: having linear dependent normal vectors). With a mobile device where the camera is very close to the screen and on the same plane, and the user's iris faces the camera/screen head on, this assumption is typically correct.

So, not sure what your setup is, but is your camera on the same plane as the screen?

When the user tilts her head, it seems that the iris plane could also tilt and then it might not be parallel to the screen/camera plane anymore. If that is the case, you could try to estimate the iris plane (by estimating its normal vector) somehow and de-tilt it, like they do with the head/eye tilt in this following publication, https://www.mpi-inf.mpg.de/departments/computer-vision-and-machine-learning/research/gaze-based-human-computer-interaction/revisiting-data-normalization-for-appearance-based-gaze-estimation/

Rasmus

[DavitMob]

Thanks for the reply Rasmus. It was very helpful.