imgaug_transform

lightning_pose.data.augmentations.imgaug_transform(params_dict: dict | DictConfig) → Sequential[source]

Create simple and flexible data transform pipeline that augments images and keypoints.

Parameters:

params_dict – each key must be the name of a transform importable from imgaug.augmenters, e.g. β€œAffine”, β€œFliplr”, etc. The value must be a dict with several optional keys: - β€œp” (float): probability of applying transform (using imgaug.augmenters.Sometimes) - β€œargs” (list): arguments for transform - β€œkwargs” (dict): keyword args for the transformation

Examples

Create a pipeline with - Affine transformation applied 50% of the time with rotation uniformly sampled from

(-25, 25) degrees

  • MotionBlur transformation that is applied 25% of the time with a kernel size of 5 pixels and blur direction uniformly sampled from (-90, 90) degrees

>>> params_dict = {
>>>    'Affine': {'p': 0.5, 'kwargs': {'rotate': (-25, 25)}},
>>>    'MotionBlur': {'p': 0.25, 'kwargs': {'k': 5, 'angle': (-90, 90)}},
>>> }

In a config file, this will look like: >>> training: >>> imgaug: >>> Affine: >>> p: 0.5 >>> kwargs: >>> rotate: [-10, 10] >>> MotionBlur: >>> p: 0.25 >>> kwargs: >>> k: 5 >>> angle: [-90, 90]

Create a pipeline with - Rot90 transformation applied 100% of the time with rotations of 0, 90, 180, 270 degrees.

>>> params_dict = {
>>>     'Rot90': {'p': 1.0, 'kwargs': {'k': [[0, 1, 2, 3]]}},  # note required nested list
>>> }

In a config file, this will look like: >>> training: >>> imgaug: >>> Rot90: >>> p: 1.0 >>> kwargs: >>> k: [0, 1, 2, 3]

NOTE: if you pass a list of exactly 2 values to Rot90 it will be parsed as a tuple and all (discrete) rotations between the two values will be sampled uniformly. For example, k: [0, 2] is equivalent to k: [0, 1, 2]. If you need to _only_ sample two non-contiguous integers please raise an issue.

Returns:

imgaug pipeline