EasyPose

Introduction

EasyPoseis a human pose estimation algorithm library based on Python and Onnxruntime, which integrates commonly used human pose estimation algorithms and can be used after installation. The original intention of developing EasyPose is to enable developers to easily use various human pose estimation algorithms for their own tasks. Therefore, EasyPose has fewer third-party dependencies and is more convenient to use. When developing a human pose estimation application using EasyPose, the program automatically downloads the corresponding weight file from the server and only requires less than ten lines of code to obtain the algorithm's prediction results.

Major features

• Simple operation

  EasyPose can quickly call various human pose estimation algorithms with simple commands and supports custom models, greatly facilitating developers to quickly use algorithms.

• Supports multiple models

  EasyPose supports various human pose estimation models, including HRNet and RTMPose, as well as human detection models such as RTMDet and YOLOv8.

• Fast speed

  The model implements the GPU version and can quickly call algorithms within the GPU.

---

Install

1、Install anaconda and create a new virtual environment

conda create -n easypose python=3.8

2、Clone the repository and install EasyPose

git clone https://github.com/Dominic23331/EasyPose.git
pip install -v -e .

3、Verify installation

import easypose as ep
print(ep.pose_model_list())

---

Models

Human detection algorithm

Model	Input Size	AP	PARAMS	GFLOPS
RTMDet-tiny	640x640	41.1	4.8	8.1
RTMDet-s	640x640	44.6	8.89	14.8
YOLOv8-n	640x640	37.3	3.2	8.7
YOLOv8-s	640x640	44.9	11.2	28.6

The above model is taken from mmdetection and ultralytics.

Human pose estimation algorithm

Model	Input Size	AP	AR
RTMPose-tiny	256x192	68.2	73.6
RTMPose-s	256x192	71.6	76.8
RTMPose-m	256x192	74.6	79.5
RTMPose-l	256x192	75.8	80.6
ResNet50-SimCC	256x192	72.1	78.1
ResNet50-Heatmap	256x192	72.0	77.5
HRNet-Heatmap	256x192	74.9	80.4
HRNet-Dark	256x192	75.7	80.7
Hourglass	256x192	72.6	78.0
Lite-HRNet-Heatmap	256x192	64.2	70.5
MobileNetv2-Heatmap	256x192	64.8	70.9
MobileNetv2-SimCC	256x192	62.0	67.8

The above model is taken from mmpose。

---

Getting Start

1、Importing the EasyPose and OpenCV libraries

import easypose as ep
import cv2

2、Instantiating the TopDown model

model = ep.TopDown('rtmpose_s', 'SimCC', 'rtmdet_s')

3、Using the predict function to predict input images

image = cv2.imread('img.jpg')
poses = model.predict(image)

4、Draw key points of the human body in the image

image = ep.draw_keypoints(image, poses)

---

Open Source License

EasyPose follows Apache 2.0 open source license.

---

Subsequent tasks

• Add more TopDown human pose estimation algorithms
• Add some one-stage human pose estimation algorithms
• Optimize the speed of existing models and add quantitative models
• Writing instructional documents
• Publish wheel files in pypi
• Support whole body pose estimation algorithms
• Support MPII dataset
• Suppoty animal pose estimation algorithms