luxonis_ml.data.augmentations.AlbumentationsEngine

Configuration Format

The configuration contains a list of transformations, each specified by its name and optional parameters as described in the AlbumentationConfigItem schema.

The name must be either a valid name of an Albumentations transformation, or a name of a custom transformation registered in the TRANSFORMATIONS registry.

For example:

[
    {
        "name": "Affine",
        "params": {
            "rotate": 30,
            "scale": 0.5,
            "p": 0.3,
        },
    },
    {
        "name": "MixUp",
        "params": {
            "alpha": [0.3, 0.7],
            "p": 0.5,
        },
    },
    {
        "name": "CustomResize",
        "use_for_resizing": True,
    },
]

Transformation Order

The order of transformations provided in the configuration is not guaranteed to be preserved. The transformations are divided into the following groups and are applied in this order:

1. Batch transformations: Subclasses of BatchTransform.
2. Spatial transformations: Subclasses of A.DualTransform.
3. Custom transformations: Subclasses of A.BasicTransform, but not subclasses of more specific base classes above.
4. Pixel transformations: Subclasses of A.ImageOnlyTransform. These transformations act only on the image.

Supported Augmentations

Augmenting Unsupported Types

Albumentations does not natively support all label types supported by Luxonis Data Format. This section describes how unsupported types are handled.

Note that the following applies only to officially supported augmentations. Custom augmentations can be implemented to handle arbitrary types.

Custom Augmentations

Custom augmentations can be implemented by creating a subclass of A.BasicTransform and registering it in the TRANSFORMATIONS registry.

Possible target types that the augmentation can receive are:

• "image":

  The image. All augmentations should usually support this target. For subclasses of A.ImageOnlyTransform or A.DualTransform this means overriding apply.

• "bboxes":

  Bounding boxes. For subclasses of A.DualTransform, this means overriding apply_to_bboxes.

• "keypoints":

  Keypoints. For subclasses of A.DualTransform, this means overriding apply_to_keypoints.

• "mask":

  Segmentation masks. For subclasses of A.DualTransform, this means overriding apply_to_mask.

• "instance_mask":

  Instance segmentation masks. For subclasses of BatchTransform, this means overriding apply_to_instance_mask.

  Subclasses of A.DualTransform do not support this target; instance masks are treated as regular masks instead.

  Custom augmentations can support instance masks by implementing their own logic for handling them and overriding the targets property to include the "instance_mask" target.

• "array":

  Arbitrary arrays. Can only be supported by custom augmentations by implementing their own logic and adding the "array" target to the targets property.

• "metadata":

  Metadata labels. Same situation as with the "array" type.

• "classification":

  One-hot encoded multi-task classification labels. Same situation as with the "array" type.

For example:

class CustomArrayAugmentation(A.BasicTransform):

    @property
    @override
    def targets(self):
        return {
            "image": self.apply,
            "array": self.apply_to_array,
        }

    @override
    def apply(self, image, **kwargs):
        ...

    def apply_to_array(
        self, array: np.ndarray, **kwargs
    ) -> np.ndarray:
        ...