PyMotion: A Python Library for Motion Data

PyMotion is a Python library that provides various functions for manipulating and processing motion data in NumPy or PyTorch. It is designed to facilitate the development of neural networks for character animation.

Some features of PyMotion are:

• A comprehensive set of quaternion operations and conversions to other rotation representations, such as rotation matrix, axis-angle, euler, and 6D representation
• A dual quaternion representation for rigid displacements, which can help neural networks better understand poses, as proposed by Andreou et al. [2022] and later adopted by Ponton et al. [2023]
• A continuous 6D rotation representation, as introduced by Zhou et al. [2019]
• A BVH file reader and preprocessor for loading and transforming motion data
• Skeletal operations such as Forward Kinematics for computing global joint positions from local joint rotations
• [Deprecated] A plotly-based visualizer for debugging and visualizing character animation directly in Python
• PyMotion to Blender automatic communication for debugging and visualizing character animation
• NumPy and PyTorch implementations and tests for all functions

Contents

1. Installation
2. Examples
3. Roadmap
4. License

Installation

1. [Optional] Install PyTorch using Pip as instructed in their webpage.

2. Install PyMotion:

pip install upc-pymotion

3. [Optional] Install Plotly and Dash for the visualizer (no needed for Blender visualization):

pip install upc-pymotion[viewer]

Examples

The examples below are small API snippets. Larger, runnable examples live in examples/; each example has its own README and optional dependencies so research-specific packages are not installed with PyMotion itself.

• examples/bones_seed_soma/ loads BONES-SEED SOMA BVH files with PyMotion and runs them through SOMA.

import numpy as np
from pymotion.io.bvh import BVH

bvh = BVH()
bvh.load("test.bvh")

print(bvh.data["names"])
# Example Output: ['Hips', 'LeftHip', 'LeftKnee', 'LeftAnkle', 'LeftToe', 'RightHip', 'RightKnee', 'RightAnkle', 'RightToe', 'Chest', 'Chest3', 'Chest4', 'Neck', 'Head', 'LeftCollar', 'LeftShoulder', 'LeftElbow', 'LeftWrist', 'RightCollar', 'RightShoulder', 'RightElbow', 'RightWrist']


# Move root joint to (0, 0, 0)
local_rotations, local_positions, parents, offsets, end_sites, end_sites_parents = bvh.get_data()
local_positions[:, 0, :] = np.zeros((local_positions.shape[0], 3))
bvh.set_data(local_rotations, local_positions)

# Scale the skeleton
bvh.set_scale(0.75)

bvh.save("test_out.bvh")

NumPy

from pymotion.io.bvh import BVH
from pymotion.ops.skeleton import fk

bvh = BVH()
bvh.load("test.bvh")

local_rotations, local_positions, parents, offsets, end_sites, end_sites_parents = bvh.get_data()
global_positions = local_positions[:, 0, :]  # root joint
pos, rotmats = fk(local_rotations, global_positions, offsets, parents)

PyTorch

from pymotion.io.bvh import BVH
from pymotion.ops.skeleton_torch import fk
import torch

bvh = BVH()
bvh.load("test.bvh")

local_rotations, local_positions, parents, offsets, end_sites, end_sites_parents = bvh.get_data()
global_positions = local_positions[:, 0, :]  # root joint
pos, rotmats = fk(
    torch.from_numpy(local_rotations),
    torch.from_numpy(global_positions),
    torch.from_numpy(offsets),
    torch.from_numpy(parents),
)

NumPy

import pymotion.rotations.quat as quat
import numpy as np

angles = np.array([np.pi / 2, np.pi, np.pi / 4])[..., np.newaxis]
# angles.shape = [3, 1]
axes = np.array([[1, 0, 0], [0, 1, 0], [0, 0, 1]])
# axes.shape = [3, 3]

q = quat.from_angle_axis(angles, axes)

rotmats = quat.to_matrix(q)

euler = quat.to_euler(q, np.array([["x", "y", "z"], ["z", "y", "x"], ["y", "z", "x"]]))
euler_degrees = np.degrees(euler)

scaled_axis = quat.to_scaled_angle_axis(q)

PyTorch

import pymotion.rotations.quat_torch as quat
import numpy as np
import torch

angles = torch.Tensor([torch.pi / 2, torch.pi, torch.pi / 4]).unsqueeze(-1)
# angles.shape = [3, 1]
axes = torch.Tensor([[1, 0, 0], [0, 1, 0], [0, 0, 1]])
# axes.shape = [3, 3]

q = quat.from_angle_axis(angles, axes)

rotmats = quat.to_matrix(q)

euler = quat.to_euler(q, np.array([["x", "y", "z"], ["z", "y", "x"], ["y", "z", "x"]]))
euler_degrees = torch.rad2deg(euler)

scaled_axis = quat.to_scaled_angle_axis(q)

NumPy

from pymotion.io.bvh import BVH
import pymotion.ops.skeleton as sk
import numpy as np

bvh = BVH()
bvh.load("test.bvh")

local_rotations, local_positions, parents, offsets, end_sites, end_sites_parents = bvh.get_data()

root_dual_quats = sk.to_root_dual_quat(
    local_rotations, local_positions[:, 0, :], parents, offsets
)

local_translations, local_rotations = sk.from_root_dual_quat(root_dual_quats, parents)
global_positions = local_translations[:, 0, :]
offsets = local_translations.copy()
offsets[:, 0, :] = np.zeros((offsets.shape[0], 3))

PyTorch

from pymotion.io.bvh import BVH
import pymotion.ops.skeleton_torch as sk
import torch

bvh = BVH()
bvh.load("test.bvh")

local_rotations, local_positions, parents, offsets, end_sites, end_sites_parents = bvh.get_data()

root_dual_quats = sk.to_root_dual_quat(
    torch.from_numpy(local_rotations),
    torch.from_numpy(local_positions[:, 0, :]),
    torch.from_numpy(parents),
    torch.from_numpy(offsets),
)

local_translations, local_rotations = sk.from_root_dual_quat(root_dual_quats, parents)
global_positions = local_translations[:, 0, :]
offsets = local_translations.clone()
offsets[:, 0, :] = torch.zeros((offsets.shape[0], 3))

NumPy

from pymotion.io.bvh import BVH
import pymotion.rotations.ortho6d as sixd

bvh = BVH()
bvh.load("test.bvh")

local_rotations, _, _, _, _, _ = bvh.get_data()

continuous = sixd.from_quat(local_rotations)

local_rotations = sixd.to_quat(continuous)

PyTorch

from pymotion.io.bvh import BVH
import pymotion.rotations.ortho6d_torch as sixd
import torch

bvh = BVH()
bvh.load("test.bvh")

local_rotations, _, _, _, _, _ = bvh.get_data()

continuous = sixd.from_quat(torch.from_numpy(local_rotations))

local_rotations = sixd.to_quat(continuous)

NumPy

import numpy as np
from pymotion.io.bvh import BVH
from pymotion.ops.skeleton import fk
from pymotion.ops.skeleton import from_root_positions

bvh = BVH()
bvh.load("test.bvh")
local_rotations, local_positions, parents, offsets, _, _ = bvh.get_data()
pos, _ = fk(local_rotations, np.zeros((local_positions.shape[0], 3)), offsets, parents)

pred_rots = from_root_positions(pos, parents, offsets)

bvh.set_data(pred_rots, local_positions)
bvh.save("test_out.bvh")  # joint positions should be similar as test.bvh

PyTorch

import torch
from pymotion.io.bvh import BVH
from pymotion.ops.skeleton_torch import fk
from pymotion.ops.skeleton_torch import from_root_positions

bvh = BVH()
bvh.load("test.bvh")
local_rotations, local_positions, parents, offsets, _, _ = bvh.get_data()
offsets = torch.from_numpy(offsets)
parents = torch.from_numpy(parents)
pos, _ = fk(
    torch.from_numpy(local_rotations),
    torch.zeros((local_positions.shape[0], 3)),
    offsets,
    parents,
)

pred_rots = from_root_positions(pos, parents, offsets)

bvh.set_data(pred_rots.numpy(), local_positions)
bvh.save("test_out.bvh")  # joint positions should be similar as test.bvh

NumPy

import numpy as np
from pymotion.io.bvh import BVH
from pymotion.ops.skeleton import mirror

bvh = BVH()
bvh.load("test.bvh")

names = bvh.data["names"].tolist()
print(names)
# ['Hips', 'Chest', 'Chest2', 'Chest3', 'Chest4', 'Neck', 'Head', 'RightCollar', 'RightShoulder',
#  'RightElbow', 'RightWrist', 'LeftCollar', 'LeftShoulder', 'LeftElbow', 'LeftWrist', 'RightHip',
#  'RightKnee', 'RightAnkle', 'RightToe', 'LeftHip', 'LeftKnee', 'LeftAnkle', 'LeftToe']

joints_mapping = np.array(
    [
        (
            names.index("Left" + n[5:])
            if n.startswith("Right")
            else (names.index("Right" + n[4:]) if n.startswith("Left") else names.index(n))
        )
        for n in names
    ]
)

local_rotations, local_positions, parents, offsets, end_sites, _ = bvh.get_data()

mirrored_local_rots, mirrored_global_pos, mirrored_offsets, _ = mirror(
    local_rotations,
    local_positions[:, 0, :],  # global position of the root joint
    parents,
    offsets,
    end_sites,
    joints_mapping=joints_mapping,  # joints_mapping is only required for mode="symmetry"
    mode="symmetry",  # other modes: "all" or "positions"
    axis="X",
)

local_positions[:, 0, :] = mirrored_global_pos
bvh.set_data(mirrored_local_rots, local_positions)
# Uncomment when mode == "all"
# bvh.data["offsets"] = mirrored_offsets
bvh.save("test_mirrored.bvh")

PyTorch

import torch
from pymotion.io.bvh import BVH
from pymotion.ops.skeleton_torch import mirror

bvh = BVH()
bvh.load("test.bvh")

names = bvh.data["names"].tolist()
print(names)
# ['Hips', 'Chest', 'Chest2', 'Chest3', 'Chest4', 'Neck', 'Head', 'RightCollar', 'RightShoulder',
#  'RightElbow', 'RightWrist', 'LeftCollar', 'LeftShoulder', 'LeftElbow', 'LeftWrist', 'RightHip',
#  'RightKnee', 'RightAnkle', 'RightToe', 'LeftHip', 'LeftKnee', 'LeftAnkle', 'LeftToe']

joints_mapping = torch.Tensor(
    [
        (
            names.index("Left" + n[5:])
            if n.startswith("Right")
            else (names.index("Right" + n[4:]) if n.startswith("Left") else names.index(n))
        )
        for n in names
    ]
).to(torch.int32)

local_rotations, local_positions, parents, offsets, end_sites, _ = bvh.get_data()

mirrored_local_rots, mirrored_global_pos, mirrored_offsets, _ = mirror(
    torch.from_numpy(local_rotations),
    torch.from_numpy(local_positions[:, 0, :]),  # global position of the root joint
    torch.from_numpy(parents),
    torch.from_numpy(offsets),
    torch.from_numpy(end_sites),
    joints_mapping=joints_mapping,  # joints_mapping is only required for mode="symmetry"
    mode="symmetry",  # other modes: "all" or "positions"
    axis="X",
)

local_positions[:, 0, :] = mirrored_global_pos.numpy()
bvh.set_data(mirrored_local_rots.numpy(), local_positions)
# Uncomment when mode == "all"
# bvh.data["offsets"] = mirrored_offsets.numpy()
bvh.save("test_mirrored.bvh")

from pymotion.render.viewer import Viewer
from pymotion.io.bvh import BVH
from pymotion.ops.skeleton import fk

bvh = BVH()
bvh.load("test.bvh")

local_rotations, local_positions, parents, offsets, _, _ = bvh.get_data()
global_positions = local_positions[:, 0, :]  # root joint
pos, rotmats = fk(local_rotations, global_positions, offsets, parents)

viewer = Viewer(use_reloader=True, xy_size=5)
viewer.add_skeleton(pos, parents)
# add additional info using add_sphere(...) and/or add_line(...), examples:
# viewer.add_sphere(sphere_pos, color="green")
# view