This repository contains the cource code for the paper HHP-Net: A light Heteroscedastic neural network for Head Pose estimation with uncertainty [WACV22].
Code Author: Giorgio Cantarini
Any questions or discussions are welcomed!
In this paper we introduce a novel method to estimate the head pose of people in single images starting from a small set of head keypoints. To this purpose, we propose a regression model that exploits keypoints and outputs the head pose represented by yaw, pitch, and roll. Our model is simple to implement and more efficient with respect to the state of the art -- faster in inference and smaller in terms of memory occupancy -- with comparable accuracy. Our method also provides a measure of the heteroscedastic uncertainties associated with the three angles, through an appropriately designed loss function. As an example application, we address social interaction analysis in images: we propose an algorithm for a quantitative estimation of the level of interaction between people, starting from their head poses and reasoning on their mutual positions.
To clone the repository:
git clone https://github.com/Malga-Vision/HHP-Net.gitTo install the requirements:
pip install -r requirements.txt
There are different choices for the key points detector: in this repository we propose two variants
- a
normalversion: accurate but less efficient - a
fasterversion: less accurate but faster
We test three different backbones of CenterNet (HourGlass104, Resnet50V2 and Resnet50V1 available in the TensorFlow 2 Detection Model Zoo); each model takes as input 512x512 images.
Download one of the previous model (e.g. HourGlass104) then extract it to HHP-Net/centernet/ with:
tar -zxvf centernet_hg104_512x512_kpts_coco17_tpu-32.tar.gz -C /HHP-Net/centernetTo make inference on a single image, run:
python inference_on_image.py [--detection-model PATH_DETECTION_MODEL] [--hhp-model PATH_HHPNET] [--image PATH_IMAGE]
To make inference on frames from the webcam, run:
python inference_on_webcam.py [--detection-model PATH_DETECTION_MODEL] [--hhp-model PATH_HHPNET]
To estimate the keypoints firstly we use an object detection model for detecting people; then we exploit a model to estimate the pose of each people detected by the previous model in the image.
In order to detect people we test Centernet MobilenetV2:
download it and then extract it to HHP-Net/centernet/:
tar -zxvf centernet_mobilenetv2fpn_512x512_coco17_od.tar.gz -C /HHP-Net/centernetThen download Posenet for pose estimation and move to HHP-Net/posenet/
mv posenet_mobilenet_v1_100_257x257_multi_kpt_stripped.tflite HHP-Net/posenet/To make inference on a single image, run:
python fast_inference_on_image.py [--detection-model PATH_MODEL_DETECTION] [--pose-model PATH_MODEL_POSE] [--hhp-model PATH_HHPNET] [--image PATH_IMAGE]
To make inference on frames from the webcam, run:
python fast_inference_on_webcam.py [--detection-model PATH_MODEL_DETECTION] [--pose-model PATH_MODEL_POSE] [--hhp-model PATH_HHPNET]
To make inference on frames from a video, run:
python fast_inference_on_webcam.py [--detection-model PATH_MODEL_DETECTION] [--pose-model PATH_MODEL_POSE] [--hhp-model PATH_HHPNET] [--video PATH_VIDEO]
If you find this code useful for your research, please use the following BibTeX entry.
@inproceedings{cantarini2022hhp,
title={HHP-Net: A light Heteroscedastic neural network for Head Pose estimation with uncertainty},
author={Cantarini, Giorgio and Tomenotti, Federico Figari and Noceti, Nicoletta and Odone, Francesca},
booktitle={Proceedings of the IEEE/CVF Winter Conference on applications of computer vision},
pages={3521--3530},
year={2022}
}
MIT