---
license: apache-2.0
task_categories:
  - tabular-classification
tags:
  - pose-estimation
  - mediapipe
  - core-exercises
  - physical-therapy
  - xgboost
  - biomechanics
  - rehabilitation
size_categories:
  - 1K<n<10K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data.csv
---

# CoreExercise5K: A Machine Learning Dataset for Core Exercise Assessment

**CoreExercise5K** is a tabular dataset composed of 3D skeletal landmark coordinates and joint angles extracted using MediaPipe Pose. It was curated for machine-learning-based physical therapy and real-time core exercise posture auditing.

---

## 📌 Dataset Overview

- **Total Samples:** 5,016 rows
- **Number of Features:** 108 columns (33 MediaPipe pose landmarks $\times$ $x, y, z$ coordinates + derived joint angle attributes)
- **Target Variable (`label`):** 9 classes (8 core exercise movements + neutral/none)
- **Modality:** Tabular / Skeletal Landmarks (Normalized 3D coordinates)

---

## 🏋️‍♂️ Class Distribution

| Exercise Class | Sample Count | Percentage |
| :--- | :--- | :--- |
| **Bird Dog** | 965 | 19.24% |
| **Plank** | 830 | 16.55% |
| **None (Neutral / Inactive)** | 622 | 12.40% |
| **High Plank** | 579 | 11.54% |
| **Dead Bug** | 492 | 9.81% |
| **Glute Bridge** | 479 | 9.55% |
| **Side Plank** | 416 | 8.29% |
| **Dart** | 357 | 7.12% |
| **Knee Plank** | 276 | 5.50% |
| **Total** | **5,016** | **100%** |

---

## 🔬 Data Collection & Preprocessing

1. **Raw Frame Acquisition:** 920 high-quality frames across diverse subjects performing physiotherapy core exercises were curated.
2. **Data Augmentation:** An augmentation pipeline applied 6 transformations to each frame:
   - Original frame
   - $+15^\circ$ rotation
   - $-15^\circ$ rotation
   - Horizontal flip
   - Flipped with $+15^\circ$ rotation
   - Flipped with $-15^\circ$ rotation  
   *(Yielding a theoretical maximum of $920 \times 6 = 5,520$ frames)*
3. **Landmark Detection & Quality Filtering:** Each augmented frame was passed through the **MediaPipe Pose** model. Frames with occlusions or missing keypoints where MediaPipe failed detection were filtered out, resulting in **5,016 clean, fully labeled keypoint vectors**.

---

## 💻 Getting Started

You can load this dataset easily with the Hugging Face `datasets` library or directly with pandas:

### Using Hugging Face Datasets

```python
from datasets import load_dataset

# Load dataset
dataset = load_dataset("xgboostgod/CoreExercise5K")
df = dataset["train"].to_pandas()

print(f"Dataset shape: {df.shape}")
print(df["label"].value_counts())
```

### Using Pandas Directly

```python
import pandas as pd

url = "https://huggingface.co/datasets/xgboostgod/CoreExercise5K/resolve/main/data.csv"
df = pd.read_csv(url)
print(df.head())
```

---

## 📜 License

This dataset is released under the [Apache 2.0 License](https://www.apache.org/licenses/LICENSE-2.0).