Iris Landmark Detection without showing the whole face

[BILLXZY1215]

MediaPipe Solution (you are using)

Facial Landmark Detection / Iris Landmark Detection

Programming language

Python

Are you willing to contribute it

Yes

Describe the feature and the current behaviour/state

It is great that we can detect eye / iris based on the facial landmark detection results. As far as my understanding, the iris landmark detection algorithm firstly runs the facial landmark results (468 points) and pointed out points that related to eyes / iris, then the cropped image serves as input to another eye refinement model, and then put together we have 478 feature points on the face.

However, I'm wondering what if we do not have a whole face (e.g. only one eye with some paddings on the face), how can I get the iris / eye landmark detection results? Because in this case facial landmark detection is not be able to use anymore as only one eye is in the image, can we open-source the [Eye Refinement Model] independently?

Will this change the current API? How?

Slightly, just an alternative I/O solution.

Who will benefit with this feature?

No response

Please specify the use cases for this feature

For some tasks like eye disease diagnose, we need images that have higher resolutions than taking pictures by mobiles that showing your whole face. Thus, we have some hardware devices that can take pictures which only contains eyes, it is quite similar compared with what you have implemented in iris landmark detection: crop image.

Any Other info

No response

[BILLXZY1215]

i.e. Is it possible to open-source this trained model? Thanks a lot.

[kuaashish]

Hello @BILLXZY1215,

Thank you for raising this request, We are forwarding this internally and based on the discussion, team will prioritise the work.

[samuelwhiskeyjohnson]

Any status update? I cant wait!

[lu-wang-g]

You can refer to the legacy MediaPipe Solutions for Iris: https://github.com/google/mediapipe/blob/master/docs/solutions/iris.md, where you can find examples and demo apps. You might need to do your own plumbing work to integrate this model into your application, since we don't offer ready-to-use pipeline just for Iris at this point. The model can be found here.