Update README.md

README.md

@@ -22,3 +22,179 @@ configs:
   - split: augmented
     path: data/augmented-*
 ---
+
+# Asian vs Western Food Image Classification Dataset
+
+## Dataset Description
+
+This dataset contains food images for binary classification between Asian and Western cuisine styles, with both original and augmented versions available for training robust computer vision models for food recognition tasks.
+
+### Dataset Summary
+
+- **Task**: Binary Food Classification (Asian vs Western)
+- **Total Images**: 360 (40 original + 320 augmented)
+- **Image Size**: 224x224 pixels
+- **Format**: RGB images
+- **Labels**: Binary (0 or 1)
+- **Course**: CMU 24-679 Homework 1
+
+## Dataset Structure
+
+### Data Splits
+
+The dataset contains two splits:
+
+| Split | Number of Images | Description |
+|-------|-----------------|-------------|
+| `original` | 40 | Original food images at 224x224 resolution |
+| `augmented` | 320 | Augmented versions of original images (8x augmentation) |
+
+### Data Fields
+
+- **`image`**: PIL Image object (224x224 RGB)
+- **`label`**: Binary integer label (0 or 1)
+  - 0: Western cuisine
+  - 1: Asian cuisine
+
+## Usage
+
+### Loading the Dataset
+
+```python
+from datasets import load_dataset
+
+# Load the entire dataset
+dataset = load_dataset("maryzhang/hw1-24679-image-dataset")
+
+# Access specific splits
+original_data = dataset['original']
+augmented_data = dataset['augmented']
+
+# Get a single example
+example = original_data[0]
+image = example['image']
+label = example['label']
+```
+
+## Data Augmentation Details
+
+The augmented split was created using the following transformations:
+
+- **Geometric Transformations**:
+  - Random Resized Crop (scale: 0.7-1.0, ratio: 0.75-1.33)
+  - Random Horizontal Flip (p=0.5)
+  - Random Vertical Flip (p=0.1)
+  - Random Rotation (±15 degrees)
+
+- **Color Transformations**:
+  - Color Jitter (brightness=0.2, contrast=0.2, saturation=0.15, hue=0.05)
+  - Random Adjust Sharpness (factor=1.5, p=0.3)
+  - Random Auto Contrast (p=0.2)
+
+- **Advanced Augmentations**:
+  - RandAugment (num_ops=2, magnitude=7)
+  - Random Erasing (p=0.2, scale=0.02-0.08)
+
+Each original image generated 7 augmented variants plus the resized original, resulting in 8 images per original sample.
+
+## Dataset Creation
+
+### Source Data
+
+The original 40 food images were captured specifically for this Asian vs Western cuisine classification task. Images were collected ensuring:
+- Clear representation of distinct cuisine styles
+- Variety in dishes, plating styles, and presentations
+- No personally identifiable information
+- Consistent image quality
+- Balanced representation between both cuisine types
+
+### Preprocessing
+
+1. All images resized to 224x224 pixels using bilinear interpolation
+2. Converted to RGB format
+3. Normalized to standard PIL Image format
+
+### Collection Process
+
+Images were collected from:
+- Personal food photography
+- Restaurant meals representing authentic cuisine styles
+- Home-cooked dishes from both culinary traditions
+
+## Considerations for Using this Data
+
+### Recommended Use Cases
+
+- Training binary food classifiers
+- Transfer learning experiments with food recognition
+- Educational projects for computer vision courses
+- Cross-cultural food analysis studies
+
+### Limitations
+
+- Small dataset size (40 original samples) - transfer learning strongly recommended
+- Binary classification only - doesn't capture cuisine diversity within each category
+- May not represent all regional variations within Asian or Western cuisine
+- Augmentation may not cover all real-world food photography variations
+
+### Best Practices
+
+1. **Use pre-trained models**: Given the small dataset size, transfer learning from ImageNet or food-specific models is recommended
+2. **Cross-validation**: Use k-fold cross-validation to maximize training data usage
+3. **Split usage strategy**: 
+   - Use `original` split for validation/testing to assess real-world performance
+   - Use `augmented` split for training to improve model robustness
+4. **Consider food-specific preprocessing**: Food images may benefit from specific color normalization
+5. **Monitor class balance**: Ensure both cuisine types are equally represented during training
+
+## Model Recommendations
+
+For best results with this small dataset:
+
+```python
+# Example: Using a pre-trained ResNet model
+import torchvision.models as models
+import torch.nn as nn
+
+# Load pre-trained model
+model = models.resnet50(pretrained=True)
+
+# Freeze early layers
+for param in model.parameters():
+    param.requires_grad = False
+
+# Replace final layer for binary classification
+model.fc = nn.Sequential(
+    nn.Linear(model.fc.in_features, 256),
+    nn.ReLU(),
+    nn.Dropout(0.5),
+    nn.Linear(256, 2)  # 2 classes: Asian vs Western
+)
+```
+
+## Citation
+
+If you use this dataset, please cite:
+
+```bibtex
+@dataset{zhang2025food,
+  author = {Mary Zhang},
+  title = {Asian vs Western Food Image Classification Dataset},
+  year = {2025},
+  publisher = {Hugging Face},
+  note = {CMU 24-679 Homework 1},
+  url = {https://huggingface.co/datasets/maryzhang/hw1-24679-image-dataset}
+}
+```
+
+## License
+
+This dataset is created for educational purposes as part of CMU 24-679 coursework.
+
+## Contact
+
+For questions or issues with this dataset, please [open an issue](https://huggingface.co/datasets/maryzhang/hw1-24679-image-dataset/discussions) on the Hugging Face repository.
+
+## Acknowledgments
+
+This dataset was created as part of the requirements for CMU 24-679. Thanks to the course instructors for the assignment guidelines and framework.