State of art the Head Pose Estimation in Tensorflow2

This repository includes:

"WHENet: Real-time Fine-Grained Estimation for Wide Range Head Pose" (BMVC 2020). adapted from the original source code.
RetinaFace: Single-stage Dense Face Localisation in the Wild adapted from https://github.com/StanislasBertrand/RetinaFace-tf2.

Install

You can install this repository with pip (requires python>=3.6);

pip install headpose-detection

pip install git+https://github.com/geekysethi/headpose_estimation#egg=headpose-detection

You can also install with the setup.py

With Face Detection

To perform detection you can simple use the following lines:

import cv2
from headpose_estimation import Headpose
headpose = Headpose()
img = cv2.imread("path_to_im.jpg")
detections,image = headpose.run(img)

This will return a list of dictionary which looks like this [{'bbox': [xmin, ymin, xmax, ymax], 'yaw': yaw_value, 'pitch': pitch_value, 'roll': roll_value}

Without Face Detection

To perform detection you can simple use the following lines:

import cv2
from headpose_estimation import Headpose
headpose = Headpose(face_detection=False)
imgcrop = cv2.imread("path_to_im.jpg")
detections,image = headpose.run(imgcrop)

In this case it will return a list of dictionary which looks like this [{'yaw': yaw_value, 'pitch': pitch_value, 'roll': roll_value}

Dependncies

EfficientNet https://github.com/qubvel/efficientnet