ATUNDA - Afrogenic dance moves

This repository contains the dataset and source code used in our paper submitted to the ACM Journal of Computing and Cultural Heritage.

The dataset of Afrogenic dance moves has been systematically collected and curated for training machine learning models for dance move classification. The dataset includes nearly 400 move sequences covering 13 distinct dance moves from African and African Diaspora performative culture.

This dataset contains motion data sequences extracted from the Atunda videos using the MediaPipe pose detection library. Each subfolder under data/ corresponds to a dance move, and within each folder there are JSON files from several performances of this dance move in the following file name convension:

• <dance type>_normalized_<performance ID>.json --- normalized landmark coordinates.
• <dance type>_world_<performance ID>.json --- landmarks in world coordinates.

---

Dataset Structure

data/
├── akwaaba/
│   ├── akwaaba_normalized_1.json
│   ├── akwaaba_world_1.json
│   ├── akwaaba_normalized_2.json
│   ├── akwaaba_world_2.json
│   └── ...
├── alanta/
│   ├── alanta_normalized_1.json
│   ├── alanta_world_2.json
│   └── ...
└── ...

---

JSON Format

Each JSON file contains the motion data from a dance move sequence:

{
  "frames": [
    {
      "nose": {"x": 0.12, "y": -0.48, "z": -0.13, "visibility": 0.98},
      "left_eye_inner": {  },
      "left_eye": {  },
      "...": "...",
      "right_foot_index": {  }
    },
    
  ]
}

• The "frames" array contains the list of frames in the motion sequence of this file.
• Each frame contains 33 landmarks.
• For each landmark:
  • x, y, z: 3D coordinates (normalized or world, depending on the file)
  • visibility: Confidence score (0.0 to 1.0)

---

Landmark Order

The dataset follows the MediaPipe Pose landmark definitions.
The 33 landmarks are:

["nose", "left_eye_inner", "left_eye", "left_eye_outer", "right_eye_inner",
 "right_eye", "right_eye_outer", "left_ear", "right_ear", "mouth_left",
 "mouth_right", "left_shoulder", "right_shoulder", "left_elbow",
 "right_elbow", "left_wrist", "right_wrist", "left_pinky", "right_pinky",
 "left_index", "right_index", "left_thumb", "right_thumb", "left_hip",
 "right_hip", "left_knee", "right_knee", "left_ankle", "right_ankle",
 "left_heel", "right_heel", "left_foot_index", "right_foot_index"]

---

How to Open the Dataset in Python

Load a Single JSON File

import json

with open("data/akwaaba/akwaaba_normalized.json", "r") as f:
    data = json.load(f)

print(len(data["frames"]))  # number of frames in the sequence

first_frame = data["frames"][0]
print(first_frame["nose"])  # coordinates of the nose landmark

Load Multiple Sequences from a Folder

import json
import glob

files = glob.glob("data/akwaaba/*.json")
for file in files:
    with open(file, "r") as f:
        data = json.load(f)
        print(f"{file} → {len(data['frames'])} frames")

---

License

The data are provided under the Atunda license. By downloading the data you accept the terms and conditions of the license.

Please note that:

• Commercial use is strictly prohibited.
• The license expires in 1 year from the time you download the data.
• You must delete the dataset when your license expires.
• You can renew the license and re-download the dataset, as some of the data may have changed. (Why? Because performers may request to remove their data, or contribute new data).