spdnet-datasets

A public shared infrastructure package for SPDNet research, providing dataset loaders, covariance estimators, and synthetic data generators.

Overview

This package extracts core dataset and data generation functionality from the monolithic spdnet-benchmarks codebase, making it available as a reusable library for the SPDNet research community.

Features

Real Dataset Loaders

Pre-built loaders for hyperspectral and motion capture datasets with SPD matrices:

• Rices90: Vietnamese rice varieties (90 classes, 256×256 covariance matrices)
• HyperLeaf: Hyperspectral barley leaf imagery (4 cultivars, 3 fertilizer levels)
• HDM05: Human motion capture dataset
• UAV-HSI-Crop: UAV hyperspectral crop classification
• GSOFF: Tree species classification from airborne hyperspectral
• Chikusei: Land cover classification (19 classes, 128 spectral bands)
• DeepHS-Fruit: Fruit classification and ripeness assessment
• Placenta: Hyperspectral tissue classification
• Kaggle Wheat: Wheat disease classification

Synthetic Data Generators

Tools for generating synthetic SPD matrix datasets for controlled experiments:

• ScaleMatrixGeneratorDiagonal: Diagonal eigenvalue control with random rotation
• ScaleMatrixGeneratorBlock: Block-structured SPD matrices
• WishartGenerator: Wishart/Inverse-Wishart distributions with random SPD scales

Covariance Estimation

Utilities for computing covariance matrices from hyperspectral images:

• Sample Covariance Matrix (SCM)
• Ledoit-Wolf shrinkage estimator
• PyTorch-compatible estimators for GPU acceleration

Installation

pip install spdnet-datasets

Or for development:

git clone https://github.com/Yet-Another-Research-Organisation/spdnet-datasets.git
cd spdnet-datasets
pip install -e ".[dev,test]"

Quick Start

Loading a Real Dataset

from spdnet_datasets import DatasetManager

# Configure dataset
config = {
    'name': 'rices90',
    'path': '/path/to/data',
    'max_classes': 10,
    'seed': 42
}

# Create dataloaders
train_loader, val_loader, test_loader, num_classes = \
    DatasetManager.create_dataloaders(config)

Generating Synthetic Data

from spdnet_datasets.synthetic import ScaleMatrixGeneratorDiagonal

# Create generator
generator = ScaleMatrixGeneratorDiagonal(matrix_size=16, n_classes=3, seed=42)

# Generate data
data, labels, info = generator.generate_data(
    n_samples_per_class=120,
    max_value=100.0,
    conditioning=100.0,
    mode='geomspace'
)

Computing Covariance Matrices

from spdnet_datasets.estimator import EstimateCovariance
import numpy as np

# Initialize estimator
estimator = EstimateCovariance(method='scm', remove_mean=True)

# Compute covariance from hyperspectral image
image = np.random.rand(100, 100, 128)  # H x W x C
cov_matrix = estimator.from_image(image)  # C x C

Package Structure

spdnet-datasets/
├── src/spdnet_datasets/
│   ├── base.py              # Base dataset class
│   ├── manager.py           # Dataset manager
│   ├── real/                # Real dataset loaders
│   ├── synthetic/           # Synthetic data generators
│   ├── estimator/           # Covariance estimators
│   └── utils/               # Preprocessing utilities
└── tests/                   # Test suite

License

MIT License - see LICENSE file for details.

Citation

If you use this package in your research, please cite:

@software{spdnet_datasets,
  title = {spdnet-datasets: Dataset loaders and generators for SPDNet research},
  author = {Gallet, Matthieu and Mian, Ammar},
  year = {2026},
  url = {https://github.com/Yet-Another-Research-Organisation/spdnet-datasets}
}

Contributing

Contributions are welcome! Please feel free to submit pull requests.

Support

For issues and questions, please use the GitHub issue tracker: https://github.com/Yet-Another-Research-Organisation/spdnet-datasets/issues