README.md

metadata

license: mit

WholeBody3D-to-AvatarPose Dataset

Overview

The WholeBody3D-to-AvatarPose Dataset is a synthetic dataset designed to learn a mapping from 3D whole-body pose estimation outputs to avatar joint rotations represented as quaternions.

The primary objective is to enable the training of models that convert COCO WholeBody (133 keypoints) 3D joint coordinates into physically plausible avatar poses, suitable for real-time animation, sign language avatars, and embodied agents.

---

Task Description

Input

• 3D joint coordinates following the COCO WholeBody 133-point model
• Coordinate format:

  [num_joints=133, 3]  → (x, y, z)
  

Output

• Avatar joint rotations expressed as quaternions
• Format:

  [num_avatar_joints, 4] → (x, y, z, w)
  

The dataset enables supervised learning for:

• WholeBody3D → Avatar pose regression
• Pose retargeting
• Conditional avatar animation

---

Data Generation

All samples were synthetically generated in Blender, ensuring:

• Noise-free 3D joint positions
• Exact ground truth joint rotations
• Controlled articulation limits and pose diversity

The full data generation and training pipeline is available here:

https://github.com/fhswf/avatar-pose

This repository includes:

• Blender-based pose generation
• COCO WholeBody joint extraction
• Quaternion export per avatar bone
• Reference training implementations

---

Pose Categories

Each sample belongs to one of the following coarse-grained pose categories:

• pointing
• fist
• open_hand
• relaxed
• grasping

These categories can be used for dataset balancing, conditional models, or evaluation.

---

Data Format

Root Structure

The JSON file has two main top-level objects:

Field	Type	Description
metadata	object	Contains general information about the export
bones	object	Dictionary of bone objects keyed by bone names

Metadata Object

Field	Type	Description
armature_name	string	Name of the armature (e.g., "KIM_caucasian_male")
export_time	string	ISO 8601 timestamp of export
blender_version	string	Version of Blender used for export
bone_count	integer	Total number of bones in the armature
end_bone_count	integer	Number of end bones (bones without children)
spaces_included	array[string]	List of coordinate spaces included in the export
note	string	Usage note (e.g., replay instructions)

Bones Object

The bones object contains a dictionary where keys are bone names and values are bone objects.

Bone Object Structure

Field	Type	Description
name	string	Name of the bone (duplicates the dictionary key)
local	object	Transform data in bone's local space
armature_space	object	Transform data in armature space
world_space	object	Transform data in world space
bone_length_armature_space	float	Length of the bone in armature space
parent	string or null	Name of parent bone (null for root)
has_children	boolean	Whether this bone has child bones
is_end_bone	boolean	Whether this bone is an end bone (no children)

Local Transform Object

Field	Type	Description
location	object	Translation vector as {x, y, z}
rotation_quaternion	object	Rotation as quaternion {w, x, y, z}
scale	object	Scale vector as {x, y, z}
matrix_basis	4x4 array	4x4 transformation matrix for pose
rotation_mode	string	Rotation representation method (always "QUATERNION")

Armature Space Object

Field	Type	Description
head	object	Start point of bone as {x, y, z}
tail	object	End point of bone as {x, y, z}

World Space Object

Field	Type	Description
head	object	Start point in world coordinates {x, y, z}
tail	object	End point in world coordinates {x, y, z}
rotation_quaternion	object	World rotation as quaternion {w, x, y, z}
matrix	4x4 array	Complete 4x4 world transformation matrix

Coordinate Spaces Explanation

1. LOCAL_CHANNELS: Bone transformations relative to its rest pose (as shown in Blender's Transform panel)
2. ARMATURE_SPACE: Bone positions relative to armature origin
3. WORLD_SPACE: Bone positions in global scene coordinates

Important Notes

• The local.matrix_basis field contains the exact pose transformation and should be used for precise pose replication
• Quaternion format uses w as the scalar component and (x, y, z) as the vector components
• All transformation matrices are 4x4 and stored in row-major order
• The spaces_included array indicates which coordinate spaces are available in the export

Example Bone Entry

"Root": {
  "name": "Root",
  "local": {
    "location": {"x": 0.0, "y": 0.0, "z": 0.0},
    "rotation_quaternion": {"w": 0.9998, "x": 0.0068, "y": -0.0153, "z": -0.0092},
    "scale": {"x": 1.0, "y": 1.0, "z": 1.0},
    "matrix_basis": [[...4x4 matrix...]],
    "rotation_mode": "QUATERNION"
  },
  "armature_space": {
    "head": {"x": 0.0, "y": 0.0707, "z": -0.0096},
    "tail": {"x": 0.0013, "y": -0.0001, "z": -0.0010}
  },
  "world_space": {
    "head": {"x": 0.0, "y": 0.1340, "z": 0.9403},
    "tail": {"x": 0.0013, "y": 0.0631, "z": 0.9489},
    "rotation_quaternion": {"w": 0.0606, "x": 0.9980, "y": 0.0082, "z": -0.0159},
    "matrix": [[...4x4 world matrix...]]
  },
  "bone_length_armature_space": 0.0714,
  "parent": null,
  "has_children": true,
  "is_end_bone": false
}

Detailed schemas and joint mappings are provided in the repository.

---

Intended Use

This dataset is intended for research and development in:

• Avatar animation
• Sign language avatar synthesis
• Embodied AI
• Motion retargeting
• Quaternion-based pose regression

---

Limitations

• Synthetic data only
• Fixed avatar skeleton
• Not intended for direct biomechanical interpretation

---

Creators and Contributors

Created by:

• Galiya Yelubayeva
• Meriam Malik
• Dirk Lewanczeck

With support from:

• Professor Dr. Christian Gawron
• Jonas D. Stephan

---

Citation

If you use this dataset, please cite:

WholeBody3D-to-AvatarPose Dataset
G. Yelubayeva, M. Malik, D. Lewanczeck,
C. Gawron, J. D. Stephan

and reference the code repository:

https://github.com/fhswf/avatar-pose

---

License

Please specify the license under which this dataset is released.