This repository aims to include a complete toolkit for working with poses. It includes a new file format with Python and Javascript readers and writers, in hope to make usage simple.

The format supports any type of poses, arbitrary number of people, and arbitrary number of frames (for videos).

The main idea is having a header with instructions on how many points exists, where, and how to connect them.

The binary spec can be found in specs/v0.1.md.

pip install pose-format

To load a .pose file, use the PoseReader class:

from pose_format.pose import Pose

buffer = open("file.pose", "rb").read()
p = Pose.read(buffer)

By default, it uses NumPy for the data, but you can also use torch and tensorflow by writing:

from pose_format.pose import Pose
from pose_format.torch.pose_body import TorchPoseBody
from pose_format.tensorflow.pose_body import TensorflowPoseBody

buffer = open("file.pose", "rb").read()

p = Pose.read(buffer, TorchPoseBody)
p = Pose.read(buffer, TensorflowPoseBody)

After creating a pose object that holds numpy data, it can also be converted to torch or tensorflow:

from pose_format.numpy import NumPyPoseBody

# create a pose object that internally stores the data as a numpy array
p = Pose.read(buffer, NumPyPoseBody)

# return stored data as a torch tensor
p.torch()

# return stored data as a tensorflow tensor
p.tensorflow()

Once poses are loaded, the library offers many ways to manipulate Pose objects.

To normalize all of the data to be in the same scale, we can normalize every pose by a constant feature of their body. For example, for people we can use the average span of their shoulders throughout the video to be a constant width.

p.normalize(p.header.normalization_info(
    p1=("pose_keypoints_2d", "RShoulder"),
    p2=("pose_keypoints_2d", "LShoulder")
))

Keypoint values can be standardized to have a mean of zero and unit variance:

p.normalize_distribution()

The default behaviour is to compute a separate mean and standard deviation for each keypoint and each dimension (usually x and y). The axis argument can be used to customize this. For instance, to compute only two global means and standard deviations for the x and y dimension:

p.normalize_distribution(axis=(0, 1, 2))

p.augment2d(rotation_std=0.2, shear_std=0.2, scale_std=0.2)

To change the frame rate of a video, using data interpolation, use the interpolate_fps method which gets a new fps and a interpolation kind.

p.interpolate_fps(24, kind='cubic')
p.interpolate_fps(24, kind='linear')

Visualize an existing pose file:

from pose_format import Pose
from pose_format.pose_visualizer import PoseVisualizer

with open("example.pose", "rb") as f:
    p = Pose.read(f.read())

v = PoseVisualizer(p)

v.save_video("example.mp4", v.draw())

Draw pose on top of video:

v.save_video("example.mp4", v.draw_on_video("background_video_path.mp4"))

Convert pose to gif to easily inspect the result in Colab:

# in a Colab notebook

from IPython.display import Image

v.save_gif("test.gif", v.draw())

display(Image(open('test.gif','rb').read()))

To load an OpenPose directory, use the load_openpose_directory utility:

from pose_format.utils.openpose import load_openpose_directory

directory = "/path/to/openpose/directory"
p = load_openpose_directory(directory, fps=24, width=1000, height=1000)

Use bazel to run tests

cd pose_format
bazel test ... --test_output=errors

Alternatively, use a different testing framework to run tests, such as pytest. To run an individual test file:

pytest pose_format/tensorflow/masked/tensor_test.py

@misc{moryossef2021pose-format, 
    title={pose-format: Library for viewing, augmenting, and handling .pose files},
    author={Moryossef, Amit and M\"{u}ller, Mathias},
    howpublished={\url{https://github.com/AmitMY/pose-format}},
    year={2021}
}