CutPasteSatSeg

Repository for the paper "Evaluating the Efficacy of Cut-and-Paste Data Augmentation in Semantic Segmentation for Satellite Imagery" - IEEE IGARSS 2024

Install

The provided example makes use of PyTorch, Rasterio and the DynamicEarthNet dataset, but can be easily extended to other libraries or datasets.

pip install -r requirements.txt

Usage

Step 1: Instance Extraction

Before using the Cut-and-Paste augmentation, you need to extract instances from your dataset. Use the generate_cap_dataset.py script for this purpose.

Step 2: Instance Pasting

Extend your existing dataset class to incorporate the Cut-and-Paste augmentation:

from cut_and_paste import CutAndPaste

class YourDataset_CAP(YourDataset):
    """
    Your dataset class with Cut-and-Paste augmentation
    """

    def __init__(self, cfg, split):
        super().__init__(cfg, split)
        self.cut_and_paste = CutAndPaste(cfg["cut_and_paste"])

    def __getitem__(self, index):
        # Load image and label as in your original dataset class
        # They should be numpy arrays before applying the Cut-and-Paste augmentation

        # Apply Cut-and-Paste augmentation
        self.cut_and_paste.paste_instances(img, mask)

        # Rest of your code
        # ...
        return img, mask

Use the extended dataset class in your training pipeline:

# In your training script
dataset = YourDataset_CAP(config, split='train')
dataloader = DataLoader(dataset, batch_size=32, shuffle=True)

Configuration

The CutAndPaste class accepts a configuration dictionary with the following keys:

root: Path to the extracted instances (output_folder from Step 1)
classes: List of class indices from which to sample
num_of_instances: Number of instances to paste per image
augment_instances: Whether to apply pre-pasting augmentations

Example configuration:

cut_and_paste_config = {
    "root": "dataset/cap_dataset",
    "classes": [0, 1, 2, 3, 4, 5],
    "num_of_instances": 100,
    "augment_instances": True
}

Cite

If you find this work useful in your research, please consider citing:

Placeholder. The bibtex will be provided when the IEEE IGARSS 2024 Proceedings will be published.

IonutMotoi/CutPasteSatSeg