---
dataset_info:
  features:
  - name: image
    dtype: image
  - name: label
    dtype:
      class_label:
        names:
          '0': Lower
          '1': Upper
  splits:
  - name: original
    num_bytes: 2575827.0
    num_examples: 32
  - name: augmented
    num_bytes: 22191126.0
    num_examples: 320
  download_size: 24632318
  dataset_size: 24766953.0
configs:
- config_name: default
  data_files:
  - split: original
    path: data/original-*
  - split: augmented
    path: data/augmented-*
---

# Gym Machines Image Dataset

## Dataset Summary
This dataset includes **30+ original, student-created images** of gym machines (objects and small arrangements/scenes) with a **binary classification target** for each image:
- `0 = Lower body machine`
- `1 = Upper body machine`

The dataset is stored on Hugging Face with two splits:
- **original**: 32 manually collected and labeled images  
- **augmented**: 320 synthetic samples generated via label-preserving transformations  

Total: **350+ images**

---

## Purpose
This dataset was created as part of a course assignment to demonstrate:
- Safe collection of original image data
- Application of augmentation techniques for dataset expansion
- Preparation and publishing of datasets to Hugging Face for reproducibility and sharing

It is intended for **educational use** in computer vision, data preprocessing, and augmentation workflows.

---

## Composition
- **Subjects**: Common gym machines (e.g., leg press, hack squat, chest press, lat pulldown).  
- **Labels**: Binary (`Lower`, `Upper`).  
- **Images**: 224×224 RGB, `.jpg` format.  
- **Counts**:  
  - Original split: 32 images  
  - Augmented split: 320 images  

---

## Data Collection
- Images were captured safely by the student, without any people or personally identifiable information (PII).  
- Only objects and gym machines were included.  
- All images were resized to **224×224 pixels**.  

---

## Preprocessing & Augmentation
### Preprocessing
- Converted to RGB  
- Resized to **224×224**

### Augmentation Techniques
Applied using PyTorch/TorchVision:
- Random horizontal flip (p=0.5)  
- Random rotation (±20°)  
- Random color jitter (brightness, contrast, saturation ±0.3)  
- Random resized crop (scale = 0.8–1.0)  
- Gaussian blur  

These transformations expanded the dataset from 32 originals to 320 augmented samples, while preserving labels.

---

## Labels
- **Binary target**:  
  - `0` → Lower body machine  
  - `1` → Upper body machine  

Labels were manually assigned by the student based on machine function.  

---

## Splits
- **original** → 32 images  
- **augmented** → 320 images  
- Published as a `DatasetDict` on Hugging Face.

---

## Intended Use & Limitations
- **Use cases**: Educational exercises in dataset handling, preprocessing, augmentation, and Hugging Face dataset publishing.  
- **Not intended for**: Medical, health, or workout guidance.  
- **Limitations**:  
  - Small dataset size → not suitable for production training  
  - Labels are simplified (`Upper` vs `Lower`) and may not capture full machine usage  

---

## Ethical Considerations
- No people or personal information were included.  
- No sensitive content.  
- Strictly object-based dataset.  

---

## License
- Released under **CC BY-NC-SA 4.0** (Attribution–NonCommercial–ShareAlike).  
- You may use and adapt for educational/research purposes with attribution.  
- Not for commercial use.  

---

## AI Usage Disclosure
- AI tools (e.g., ChatGPT) assisted in:  
  - Structuring the dataset card  
- All images were **student-created**, not AI-generated.