youngwoo-yoon/PyMO

A library for machine learning research on motion capture data

PyMO

This forked version is for GENEA Challenge 2022 dataset. Code is modified to be compatible with the joint names in the dataset.

A library for using motion capture data for machine learning.

This library is currently highly experimental and everything is subject to change :)

Roadmap

• Mocap Data Parsers and Writers
• Common mocap pre-processing algorithms
• Feature extraction library
• Visualization tools

Current Features

• Read BVH Files
• Write BVH Files
• Pre-processing pipelines
  • Supporting scikit-learn API
  • Convert data representations
    • Euler angles to positions
    • Euler angles to exponential maps
    • Exponential maps to euler angles
  • Body-oriented global translation and rotation calculation with inverse tranform
  • Root-centric position normalizer with inverse tranform
  • Standard scaler
  • Joint selectors
• Visualization tools
  • Skeleton hierarchy
  • 2D frame visualization
  • 3D webgl-based animation
• Annotations
  • Foot/ground contact detector

Dependencies

numpy scipy joblib scikit-learn pandas jupyter matplotlib

Demos

See the demo notebooks for processing and visualization examples:

• bvh2features: downsampling, root-centric transformation, joint-selection and data-augmentation (mirroring)
• features2bvh: converting features back to bvh format
• Position: bvh to joint positions
• PlayMocap: Visualize in HTML

Read BVH Files

from pymo.parsers import BVHParser
import pymo.viz_tools

parser = BVHParser()

parsed_data = parser.parse('data/AV_8Walk_Meredith_HVHA_Rep1.bvh')

Get Skeleton Info

import pymo.viz_tools

viz_tools.print_skel(parsed_data)

Will print the skeleton hierarchy:

- Hips (None)
| | - RightUpLeg (Hips)
| | - RightLeg (RightUpLeg)
| | - RightFoot (RightLeg)
| | - RightToeBase (RightFoot)
| | - RightToeBase_Nub (RightToeBase)
| - LeftUpLeg (Hips)
| - LeftLeg (LeftUpLeg)
| - LeftFoot (LeftLeg)
| - LeftToeBase (LeftFoot)
| - LeftToeBase_Nub (LeftToeBase)
- Spine (Hips)
| | - RightShoulder (Spine)
| | - RightArm (RightShoulder)
| | - RightForeArm (RightArm)
| | - RightHand (RightForeArm)
| | | - RightHand_End (RightHand)
| | | - RightHand_End_Nub (RightHand_End)
| | - RightHandThumb1 (RightHand)
| | - RightHandThumb1_Nub (RightHandThumb1)
| - LeftShoulder (Spine)
| - LeftArm (LeftShoulder)
| - LeftForeArm (LeftArm)
| - LeftHand (LeftForeArm)
| | - LeftHand_End (LeftHand)
| | - LeftHand_End_Nub (LeftHand_End)
| - LeftHandThumb1 (LeftHand)
| - LeftHandThumb1_Nub (LeftHandThumb1)
- Head (Spine)
- Head_Nub (Head)

scikit-learn Pipeline API

data_pipe = Pipeline([
    ('rcpn', RootCentricPositionNormalizer()),
    ('delta', RootTransformer('abdolute_translation_deltas')),
    ('const', ConstantsRemover()),
    ('np', Numpyfier()),
    ('down', DownSampler(2)),
    ('stdscale', ListStandardScaler())
])

piped_data = data_pipe.fit_transform(train_X)

Convert to Positions

mp = MocapParameterizer('positions')

positions = mp.fit_transform([parsed_data])

Visualize a single 2D Frame

draw_stickfigure(positions[0], frame=10)

Animate in 3D (inside a Jupyter Notebook)

nb_play_mocap(positions[0], 'pos', 
              scale=2, camera_z=800, frame_time=1/120, 
              base_url='../pymo/mocapplayer/playBuffer.html')

Foot/Ground Contact Detector

from pymo.features import *

plot_foot_up_down(positions[0], 'RightFoot_Yposition')

signal = create_foot_contact_signal(pos_data[3], 'RightFoot_Yposition')
plt.figure(figsize=(12,5))
plt.plot(signal, 'r')
plt.plot(pos_data[3].values['RightFoot_Yposition'].values, 'g')

Feedback, Bugs, and Questions

For any questions, feedback, and bug reports, please use the Github Issues.

Credits

Created by Omid Alemi

Modified by Simon Alexanderson

License

This code is available under the MIT license.