DeepFashion2 Dataset Integration

This document describes the integration of the DeepFashion2 dataset with the Vestiq fashion analysis system.

Overview

DeepFashion2 is a comprehensive fashion dataset that provides:

491K diverse images of 13 popular clothing categories
Bounding box annotations for fashion items
Dense pose estimation
Commercial-consumer clothes correspondence
Scale, occlusion, zoom-in, and viewpoint labels

Integration Features

1. Dataset Loading and Processing

DeepFashion2Dataset: PyTorch dataset class for loading images and annotations
Category Mapping: Maps DeepFashion2 categories to yainage90 model categories
Data Transforms: Standard preprocessing for fashion images
Batch Processing: Efficient DataLoader implementation

2. Evaluation Framework

Detection Accuracy: Evaluate fashion object detection performance
Feature Quality: Assess feature extraction capabilities
Classification Metrics: Precision, recall, F1-score, confusion matrix
Visualization: Confusion matrix plots and performance charts

3. API Endpoints

/deepfashion2/status - Check integration status and dataset availability
/deepfashion2/statistics - Get dataset statistics and category distribution
/deepfashion2/evaluate - Run evaluation using DeepFashion2 as benchmark
/deepfashion2/setup-instructions - Get setup instructions for the dataset

Category Mapping

DeepFashion2 uses 13 detailed categories that are mapped to yainage90's 7 categories:

DeepFashion2 Category	yainage90 Category
short_sleeved_shirt	top
long_sleeved_shirt	top
short_sleeved_outwear	outer
long_sleeved_outwear	outer
vest	top
sling	top
shorts	bottom
trousers	bottom
skirt	bottom
short_sleeved_dress	dress
long_sleeved_dress	dress
vest_dress	dress
sling_dress	dress

Setup Instructions

1. Download the Dataset

The DeepFashion2 dataset requires manual download due to licensing requirements:

Visit the official repository: https://github.com/switchablenorms/DeepFashion2
Follow the dataset download instructions
Register and download the dataset files

2. Dataset Structure

Extract the dataset to ./data/deepfashion2/ with the following structure:

deepfashion2/
├── train/
│   ├── image/          # Training images
│   └── annos/          # Training annotations (JSON)
├── validation/
│   ├── image/          # Validation images
│   └── annos/          # Validation annotations (JSON)
└── test/
    ├── image/          # Test images
    └── annos/          # Test annotations (JSON)

3. Install Dependencies

Install additional dependencies for evaluation:

pip install scikit-learn matplotlib seaborn

4. Verify Setup

Check the integration status:

curl http://localhost:7861/deepfashion2/status

Usage Examples

1. Basic Dataset Loading

from deepfashion2_utils import DeepFashion2Config, DeepFashion2Dataset

config = DeepFashion2Config()
dataset = DeepFashion2Dataset(
    root_dir=config.dataset_root,
    split='validation',
    load_annotations=True
)

# Get a sample
sample = dataset[0]
print(f"Image: {sample['image_path']}")
print(f"Categories: {dataset.get_categories_in_image(sample['annotations'])}")

2. Running Evaluation

from deepfashion2_evaluation import run_full_evaluation
from fast import analyzer

# Run evaluation with 100 samples
report_path = run_full_evaluation(analyzer, max_samples=100)
print(f"Evaluation report saved to: {report_path}")

3. API Usage

# Check status
curl -X GET "http://localhost:7861/deepfashion2/status"

# Get dataset statistics
curl -X GET "http://localhost:7861/deepfashion2/statistics"

# Run evaluation
curl -X POST "http://localhost:7861/deepfashion2/evaluate?max_samples=50"

# Get setup instructions
curl -X GET "http://localhost:7861/deepfashion2/setup-instructions"

Evaluation Metrics

Detection Accuracy

Category-level accuracy: How well the model detects clothing categories
Detection score: IoU-like metric for category overlap
Confusion matrix: Detailed breakdown of predictions vs ground truth

Feature Quality

Feature dimension: Dimensionality of extracted features
Intra-category similarity: How similar features are within the same category
Inter-category distance: How well features separate different categories
Feature separability: Overall quality metric for feature discrimination

Configuration Options

DeepFashion2Config

@dataclass
class DeepFashion2Config:
    dataset_root: str = "./data/deepfashion2"
    categories: List[str] = None  # Auto-populated with 13 categories
    image_size: Tuple[int, int] = (224, 224)
    batch_size: int = 32
    num_workers: int = 4

Customization

You can customize the configuration for your specific needs:

config = DeepFashion2Config(
    dataset_root="/path/to/your/deepfashion2",
    image_size=(256, 256),
    batch_size=16
)

Performance Considerations

Memory Usage

The dataset is large (~15GB), ensure sufficient disk space
Use appropriate batch sizes based on available GPU memory
Consider using num_workers for faster data loading

CPU Optimization

The system automatically detects CPU vs GPU and optimizes accordingly
CPU inference uses float32 precision and limited threads
GPU inference uses float16 precision for better performance

Evaluation Speed

Limit max_samples for faster evaluation during development
Full evaluation on the entire validation set may take significant time
Consider running evaluations on a subset for quick feedback

Troubleshooting

Common Issues

Dataset not found: Ensure the dataset is extracted to the correct path
Permission errors: Check file permissions for the dataset directory
Memory errors: Reduce batch size or number of workers
Import errors: Install missing dependencies (scikit-learn, matplotlib, seaborn)

Debug Mode

Enable debug logging to troubleshoot issues:

import logging
logging.basicConfig(level=logging.DEBUG)

Future Enhancements

Planned Features

Training Pipeline: Fine-tune models on DeepFashion2 data
Advanced Metrics: Add more sophisticated evaluation metrics
Visualization Tools: Enhanced plotting and analysis tools
Benchmark Comparisons: Compare against other fashion datasets

Contributing

To contribute to the DeepFashion2 integration:

Fork the repository
Create a feature branch
Add tests for new functionality
Submit a pull request

References

License

This integration follows the same license as the main Vestiq project. The DeepFashion2 dataset has its own licensing terms that must be respected.