MediaPipe Iris Pytorch

Introduction

This repository demonstrate Google's MediaPipe Iris model using Python and Pytorch.

Official Project Page: https://google.github.io/mediapipe/solutions/iris.html

Iris Prediction

MediaPipe Iris Detection process consists of two parts: face landmarks detection and iris prediction with respect to eye-cropped images.

Face Landmark Model

Face landmarks are detected using the Face Landmark Model, which is the same as the MediaPipe Face Mesh model. The model predicts each face's bounding boxes and 468 3D face landmarks in real-time. The definitions of the landmark points can be found here.

Iris Model

Preprocess

Iris model takes in a 64x64 center-cropped RGB left-eye image with a 25% margin on each side.

Therefore, after obtaining the face landmarks, we have to extract eye region out for both eyes (rightEyeUpper0, rightEyeLower0, leftEyeUpper0, leftEyeLower0), see here. The center of each pupil is determined by the average of the region.

The 25% margin on each side is achieved by first determine the cropped size($d_{crop}$) using the distance between the left-most and right-most points, i.e. the eye width($d_{ew}$), for each eye region respectively. $$d_{crop} = 2 \times d_{ew}$$

We center-cropped the image using the pupil center and the calculated cropped size, then resize the image to 64x64. We also need to flip the image horizontally if it is a right eye.

Main model

The model is based on MobileNetV2 modified with customized blocks for real-time performance. It contains two outputs: eyeContours(213) and iris(15).

• Eyecontours: the eye with eyebrow region(71 points in 3D coordinates)
• Iris: pupil center(1 point in 3D) + iris contour(4 points in 3D)

Google provides the tflite model in the official project page. In this repository, I converted the model from tflite to the structure of pytorch.

More details:

• Model Card
• Paper

Get Started

Setup the environment

git clone https://github.com/Morris88826/MediaPipe_Iris.git
cd MediaPipe_Iris

conda create --name mediapipe_iris python=3.7
conda activate mediapipe_iris
pip install -r requirements.txt

Extract weights from tflite model

Run extract_iris_landmark_model.py file to extract weights in the tflite model. The extracted weights will be stored in the data folder. One can verify the result by running libs/iris.py. This will save the checkpoint file in pytorch format(.pth) if not yet created in the data folder.

python extract_iris_landmark_model.py 
python libs/iris.py

Test the result

python main.py # this will run the demo
python main.py --source {path_to_the_video}