segpaste 0.2.0

A Python package for copy-paste functionality. Torch only reimplementation of https://arxiv.org/abs/2012.07177

segpaste

A PyTorch implementation of "Simple Copy-Paste is a Strong Data Augmentation Method for Instance Segmentation" (arXiv:2012.07177).

This package also provides integration with torchvision ecosystem.

Installation

pip install segpaste

Or install from source:

git clone https://github.com/NoeFontana/segpaste.git
cd segpaste
pip install -e .

Usage

TorchVision Integration

Convenience types and wrappers are provided to ease integration with torchvision datasets and transforms.

• PaddingMask: A TVTensor representing a padding mask.
• CocoDetectionV2: A CocoDetection dataset that presents an interface compatible with torchvision.transforms.v2 and with support for padding masks.
• SanitizeBoundingBoxes: A small wrapper around torchvision.transforms.v2.SanitizeBoundingBoxes that adds support for PaddingMask.

Collator

The recommended interface is through the CopyPasteCollator which can be used in lieu of a standard collate function in a PyTorch DataLoader as long as the batch_size is greater than 1.

from segpaste import CopyPasteAugmentation, CopyPasteCollator, CopyPasteConfig

config = CopyPasteConfig()
augmentation = CopyPasteAugmentation(config)
collate_fn = CopyPasteCollator(augmentation)
torch.utils.data.DataLoader(
        dataset, # Your dataset here
        batch_size=batch_size, # Must be > 1
        collate_fn=collate_fn,
    )

Examples of usage can be found in the test suite.

Transform (Unstable API)

A minimal working example with CopyPasteTransform can be found in the examples.

Further

The public API is exposed in the segpaste namespace. It is subject to breaking changes, without prior notice, until version 1.0.0.

Development

This project uses Ruff for linting/formatting and mypy for type checking.

# Install development dependencies
pip install -e ".[dev,coco]"

# Format, lint, and type-check
ruff format . && ruff check --fix . && mypy .

# Run tests
pytest

Contributing

Contributions are welcome. Please open an issue to discuss major changes or submit a pull request.

Citation

If you use this implementation in your research, please consider citing the original paper:

@article{ghiasi2020simple,
  title={Simple Copy-Paste is a Strong Data Augmentation Method for Instance Segmentation},
  author={Ghiasi, Golnaz and Cui, Yin and Srinivas, Aravind and Qian, Rui and Lin, Tsung-Yi and Cubuk, Ekin D and Le, Quoc V and Zoph, Barret},
  journal={arXiv preprint arXiv:2012.07177},
  year={2020}
}

License

This project is licensed under the MIT License.