pybvh 0.5.0

BVH reader/writer and kinematics utilities

pybvh

A lightweight Python library for reading, writing, and manipulating BVH motion capture files. Built for researchers and developers working with skeletal animation and motion data.

Features

• Read & write BVH files with full hierarchy and motion data preservation
• Rotation conversions between Euler angles, rotation matrices, quaternions, 6D (Zhou et al.), and axis-angle — all vectorized with NumPy
• Forward kinematics to compute 3D joint positions from angles
• Skeleton operations: retargeting, scaling, joint extraction, Euler order changes
• Frame operations: slicing, concatenation, resampling to different frame rates
• Spatial transforms: mirroring, vertical rotation, speed perturbation, joint noise, root translation, frame dropout — all with seeded randomization
• Motion analysis: joint velocities/accelerations, root-relative positions, foot contact detection, normalization utilities, and a one-stop to_feature_array() export
• Batch loading of entire directories with optional parallel I/O
• NumPy export in any rotation representation — ready for any downstream workflow
• Pandas ready via an export option ready to become a DataFrame
• 3D visualization with multiple backends (matplotlib, OpenCV, k3d, vedo)

Philosophy

pybvh is framework-agnostic and outputs pure NumPy arrays. It understands motion capture data but does not assume what you'll do with it — the same library serves ML researchers, biomechanics scientists, and game developers. For ML-specific features (tensor packing, PyTorch Datasets, augmentation pipelines), see the companion library pybvh-ml.

Installation

pip install pybvh

Quick Start

import pybvh

# Load a BVH file
bvh = pybvh.read_bvh_file("walk.bvh")
print(bvh)  # 24 joints, 120 frames at 0.008333Hz

# Access motion data as NumPy arrays
bvh.root_pos          # (F, 3) root translation per frame
bvh.joint_angles      # (F, J, 3) Euler angles in degrees
bvh.joint_names       # ['Hips', 'Spine', ...] (excludes end sites)

# Get 3D joint positions via forward kinematics
coords = bvh.get_spatial_coord()  # (F, N, 3)

# Convert to other rotation representations
root_pos, quats, joints = bvh.get_frames_as_quaternion()  # (F,3), (F,J,4), joints
root_pos, rot6d, joints = bvh.get_frames_as_6d()          # (F,3), (F,J,6), joints

# Write back to file
bvh.to_bvh_file("output.bvh")

Batch Loading

Load an entire directory of BVH files and convert to NumPy arrays in one call:

from pybvh import read_bvh_directory, batch_to_numpy

# Load all BVH files from a directory
clips = read_bvh_directory("dataset/", parallel=True)

# Convert to padded NumPy array
data = batch_to_numpy(clips, representation="6d", pad=True)
# shape: (batch, max_frames, features)

Supported representations: "euler", "quaternion", "6d", "axisangle", "rotmat".

Motion Analysis

Compute motion derivatives, foot contacts, and export everything in a single array:

# Joint velocities and accelerations (finite differences of FK positions)
vel = bvh.get_joint_velocities()        # (F-1, N, 3) in units/second
acc = bvh.get_joint_accelerations()     # (F-2, N, 3)
ang_vel = bvh.get_angular_velocities()  # (F-1, J, 3) in radians/second

# Root-relative positions and trajectory
rel_pos = bvh.get_root_relative_positions()  # (F, N, 3)
traj = bvh.get_root_trajectory()             # (F, 4) ground pos + heading

# Foot contact detection (auto-detects foot joints)
contacts = bvh.get_foot_contacts()  # (F, num_feet) binary indicators

# One-stop export — flat feature array
features = bvh.to_feature_array(
    representation="6d",
    include_velocities=True,
    include_foot_contacts=True,
)  # (F-1, D)

Normalize across a dataset:

from pybvh import compute_normalization_stats, normalize_array, denormalize_array

stats = compute_normalization_stats(clips, representation="6d")
normalized = normalize_array(data, stats)
# stats are plain dicts — save with np.savez("stats.npz", **stats)

Spatial Transforms

Standard motion transforms — all support seeded randomization for reproducibility:

from pybvh import transforms

# Left-right mirroring (auto-detects joint pairs and lateral axis)
bvh_mirrored = transforms.mirror(bvh)

# Vertical rotation (auto-detects up axis)
bvh_rotated = transforms.rotate_vertical(bvh, angle_deg=90)
bvh_rotated = transforms.random_rotate_vertical(bvh, rng=np.random.default_rng(42))

# Speed perturbation (factor > 1 = faster, < 1 = slower)
bvh_fast = transforms.speed_perturbation(bvh, factor=1.5)

# Joint noise injection
bvh_noisy = transforms.add_joint_noise(bvh, sigma_deg=1.0, sigma_pos=0.5, rng=rng)

# Root translation
bvh_shifted = transforms.translate_root(bvh, offset=[100, 0, 0])

# Frame dropout with SLERP interpolation (same frame count)
bvh_dropped = transforms.dropout_frames(bvh, drop_rate=0.1, rng=rng)

All transforms also available as Bvh methods: bvh.mirror(), bvh.rotate_vertical(90), etc.

Skeleton Operations

# Change Euler rotation order for all joints
bvh_xyz = bvh.change_all_euler_orders("XYZ")

# Scale the skeleton
bvh_scaled = bvh.scale_skeleton(0.01)  # meters to centimeters

# Retarget motion to a different skeleton
bvh_retarget = bvh.change_skeleton(reference_bvh)

# Extract a subset of joints
bvh_upper = bvh.extract_joints(["Hips", "Spine", "Neck", "Head"])

# Slice and concatenate frames
clip = bvh.slice_frames(10, 50)
combined = bvh.concat(other_bvh)

# Resample to a different frame rate
bvh_30fps = bvh.resample(30)

Rotation Utilities

All functions are batch-vectorized and work on arbitrary batch dimensions:

from pybvh import rotations

# Convert between any pair of representations
R = rotations.euler_to_rotmat(angles, order="ZYX", degrees=True)
q = rotations.rotmat_to_quat(R)
aa = rotations.quat_to_euler(q, order="ZYX", degrees=True)

# Quaternion SLERP interpolation
q_mid = rotations.quat_slerp(q1, q2, t=0.5)

Visualization

from pybvh import plot

# Rest pose (T-pose)
plot.rest_pose(bvh)

# Static frame with camera control
plot.frame(bvh, frame=0, camera="front")  # also "side", "top", (azim, elev)

# Fast video export (OpenCV if installed, else matplotlib)
plot.render(bvh, "walk.mp4")

# Interactive playback (auto-detects best backend)
plot.play(bvh)

# Side-by-side comparison (sync="pad" continues to longest clip)
plot.frame([bvh1, bvh2], frame=0, labels=["Original", "Generated"])
plot.render([bvh1, bvh2], "compare.mp4", sync="pad")

# 2D root trajectory
plot.trajectory(bvh)

Install optional visualization backends for best performance:

pip install pybvh[opencv]       # Fast video rendering 
pip install pybvh[interactive]  # k3d for Jupyter notebooks
pip install pybvh[viewer]       # vedo for desktop interactive viewer
pip install pybvh[all-viz]      # All of the above

Pandas Integration

import pandas as pd

# BVH to DataFrame
df = pd.DataFrame(bvh.get_df_constructor(mode="euler"))

# DataFrame back to BVH
from pybvh import df_to_bvh
bvh_from_df = df_to_bvh(bvh.hierarchy_info_as_dict(), df)

Tutorials

The repository includes Jupyter notebooks with detailed walkthroughs:

1. Introduction to pybvh — reading, writing, and basic operations
2. Spatial coordinates — forward kinematics and 3D positions
3. Rotations — rotation representations and conversions

Requirements

• Python >= 3.9
• NumPy >= 1.21
• Matplotlib >= 3.7

Pandas is optional (pip install "pybvh[pandas]") - only used in the tutorials, not part of pybvh library.

License

MIT