Transform YAML References

Available Pre-processing Transforms

BinarizeLabelValues

• type: object

• category: Labels only

• description:

  Transforms the labels to have only 1 and 0 values. If image labels are a list (in a multi-label case) then the behaviour is to binarize with [] and [1]. For bboxes_labels, all of them will be converted to 0, since there's no "background" bbox.

• properties:

  • always_apply:

    • type: boolean
    • default: False

  • p:

    • type: number
    • default: 1.0
    • minimum: 0
    • maximum: 1

Blur

• type: object

• category: Pixel level

• description:

  Blur the input image using a random-size kernel.

• properties:

  • lower_limit:

    • type: integer
    • default: 3
    • minimum: 3
    • maximum: 100

  • upper_limit:

    • type: integer
    • default: 7
    • minimum: 5
    • maximum: 100

  • always_apply:

    • type: boolean
    • default: False

  • p:

    • type: number
    • default: 0.5
    • minimum: 0
    • maximum: 1

Cast

• type: object

• category: Others

• description:

  Casts a given set of keys to the desired type.

• properties:

  • image:

    • type: string
    • default: uint8
    • enum: ['bool', 'uint8', 'uint16', 'int8', 'uint32', 'int16', 'int32', 'int64', 'float32', 'float64']

  • label:

    • type: string
    • default: int32
    • enum: ['bool', 'uint8', 'uint16', 'int8', 'uint32', 'int16', 'int32', 'int64', 'float32', 'float64']

  • mask:

    • type: string
    • default: int32
    • enum: ['bool', 'uint8', 'uint16', 'int8', 'uint32', 'int16', 'int32', 'int64', 'float32', 'float64']

  • bboxes:

    • type: string
    • default: int32
    • enum: ['bool', 'uint8', 'uint16', 'int8', 'uint32', 'int16', 'int32', 'int64', 'float32', 'float64']

  • bboxes_labels:

    • type: string
    • default: int32
    • enum: ['bool', 'uint8', 'uint16', 'int8', 'uint32', 'int16', 'int32', 'int64', 'float32', 'float64']

  • always_apply:

    • type: boolean
    • default: False

  • p:

    • type: number
    • default: 1.0
    • minimum: 0
    • maximum: 1

ClassnameToIndex

• type: object

• category: Labels only

• description:

  Transform the string labels to indices by applying a mapping between class names and class indices.

• properties:

  • label_map:

    • type: object

  • always_apply:

    • type: boolean
    • default: False

  • p:

    • type: number
    • default: 1.0
    • minimum: 0
    • maximum: 1

Compose

• type: object

• category: Compose

• description:

  Compose a sequence of transforms.

• properties:

  • transforms:

    • type: array
    • items:
      • type: object

  • bbox_params:

    • anyOf:

      • type: null

      • type: object

  • keypoint_params:

    • anyOf:

      • type: null

      • type: object

  • additional_targets:

    • anyOf:

      • type: null

      • type: object

  • p:

    • type: number
    • default: 1.0
    • minimum: 0
    • maximum: 1

ConvertColor

• type: object

• category: Pixel level

• description:

  Converts an image to another color spaces.

• properties:

  • color_mode:

    • type: string
    • default: None
    • enum: ['RGB2GRAY', 'RGB2HSV', 'GRAY2RGB', 'RGB2LAB', 'RGB2BGR', 'RGBA2RGB']

  • always_apply:

    • type: boolean
    • default: False

  • p:

    • type: number
    • default: 1.0
    • minimum: 0
    • maximum: 1

Crop

• type: object

• category: Spatial level

• description:

  Crop a rectangular region from image.

• properties:

  • x_min:

    • type: integer
    • default: None
    • minimum: 0

  • y_min:

    • type: integer
    • default: None
    • minimum: 0

  • x_max:

    • type: integer
    • default: None
    • minimum: 0

  • y_max:

    • type: integer
    • default: None
    • minimum: 0

  • always_apply:

    • type: boolean
    • default: False

  • p:

    • type: number
    • default: 1.0
    • minimum: 0
    • maximum: 1

Downscale

• type: object

• category: Pixel level

• description:

  Decreases image quality by downscaling and upscaling back.

• properties:

  • always_apply:

    • type: boolean
    • default: False

  • p:

    • type: number
    • default: 0.5
    • minimum: 0
    • maximum: 1

ExpandBoxes

• type: object

• category: Labels only

• description:

  This transform changes the shape the bounding boxes by a certain amount, attempting that the object remains centered.

• properties:

  • height_delta:

    • type: integer
    • default: None
    • minimum: 0

  • width_delta:

    • type: integer
    • default: None
    • minimum: 0

  • always_apply:

    • type: boolean
    • default: False

  • p:

    • type: number
    • default: 1.0
    • minimum: 0
    • maximum: 1

ExpandBoxesTo

• type: object

• category: Labels only

• description:

  This transform changes the shape the bounding boxes to a predefined width and height, attempting that the object remains centered.

• properties:

  • height:

    • type: integer
    • default: None
    • minimum: 0

  • width:

    • type: integer
    • default: None
    • minimum: 0

  • always_apply:

    • type: boolean
    • default: False

  • p:

    • type: number
    • default: 1.0
    • minimum: 0
    • maximum: 1

ExpandBoxesToFit

• type: object

• category: Labels only

• description:

  Preprocessing transforms to expand bounding boxes with extreme ratios (h/w and w/h) or size smaller than a minimum size, so that it is easier to fit them with anchors in an object detection task.

• properties:

  • min_size:

    • type: integer
    • default: None
    • minimum: 0

  • max_aspect_ratio:

    • type: number
    • default: 6
    • minimum: 0

  • always_apply:

    • type: boolean
    • default: False

  • p:

    • type: number
    • default: 1.0
    • minimum: 0
    • maximum: 1

GaussianBlur

• type: object

• category: Pixel level

• description:

  Blur the input image using a Gaussian filter with a random kernel size.

• properties:

  • lower_limit:

    • type: integer
    • default: 3
    • minimum: 3
    • maximum: 100

  • upper_limit:

    • type: integer
    • default: 7
    • minimum: 5
    • maximum: 100

  • p:

    • type: number
    • default: 0.5
    • minimum: 0
    • maximum: 1

HorizontalFlip

• type: object

• category: Spatial level

• description:

  Flips the input horizontally.

• properties:

  • p:
    • type: number
    • default: 0.5
    • minimum: 0
    • maximum: 1

HueSaturationValue

• type: object

• category: Pixel level

• description:

  Randomly change hue, saturation and value of the input image.

• properties:

  • hue_shift_lower_limit:

    • type: integer
    • default: -20
    • minimum: -255
    • maximum: 255

  • hue_shift_upper_limit:

    • type: integer
    • default: 20
    • minimum: -255
    • maximum: 255

  • sat_shift_lower_limit:

    • type: integer
    • default: -30
    • minimum: -255
    • maximum: 255

  • sat_shift_upper_limit:

    • type: integer
    • default: 30
    • minimum: -255
    • maximum: 255

  • val_shift_lower_limit:

    • type: integer
    • default: -20
    • minimum: -255
    • maximum: 255

  • val_shift_upper_limit:

    • type: integer
    • default: 20
    • minimum: -255
    • maximum: 255

  • p:

    • type: number
    • default: 0.5
    • minimum: 0
    • maximum: 1

IgnoreLabelValues

• type: object

• category: Labels only

• description:

  Transforms the labels to ignore some values. In the case of masks, the ignored values are turned to zero. For bboxes, they're removed. For image labels, if label is a list they're removed, if it's a single label it's turned to zero.

• properties:

  • to_ignore:

    • type: array
    • items:
      • type: integer
      • minimum: 1

  • always_apply:

    • type: boolean
    • default: False

  • p:

    • type: number
    • default: 1.0
    • minimum: 0
    • maximum: 1

IndexToClassname

• type: object

• category: Labels only

• description:

  Transform the int labels to strings by applying a mapping between class indices and class names.

• properties:

  • label_map:

    • type: object

  • always_apply:

    • type: boolean
    • default: False

  • p:

    • type: number
    • default: 1.0
    • minimum: 0
    • maximum: 1

MotionBlur

• type: object

• category: Pixel level

• description:

  Apply motion blur to the input image using a random-size kernel.

• properties:

  • lower_limit:

    • type: integer
    • default: 3
    • minimum: 3
    • maximum: 100

  • upper_limit:

    • type: integer
    • default: 7
    • minimum: 5
    • maximum: 100

  • p:

    • type: number
    • default: 0.5
    • minimum: 0
    • maximum: 1

Normalize

• type: object

• category: Pixel level

• description:

  Divide pixel values by 255, subtract mean per channel and divide by std per channel. You images needs to be in RGB format.

• properties:

  • mean:

    • type: array
    • default: (0.485, 0.456, 0.406)
    • items:
      • type: number
      • exclusiveMinimum: 0
      • exclusiveMaximum: 1

  • std:

    • type: array
    • default: (0.229, 0.224, 0.225)
    • items:
      • type: number
      • exclusiveMinimum: 0
      • exclusiveMaximum: 1

  • always_apply:

    • type: boolean
    • default: False

  • p:

    • type: number
    • default: 1.0
    • minimum: 0
    • maximum: 1

NormalizeRange

• type: object

• category: Pixel level

• description:

  Normalizes the image such that it's values go from being in the range input_range to the range [output_range]. For example, usually the input_range is [0, 255], and with this transform you can convert the image to the range [-1, 1].

• properties:

  • output_range:

    • type: array
    • items:
      • type: number

  • input_range:

    • type: array
    • default: (0, 255)
    • items:
      • type: number

  • always_apply:

    • type: boolean
    • default: False

  • p:

    • type: number
    • default: 1.0
    • minimum: 0
    • maximum: 1

OneHot

• type: object

• category: Labels only

• description:

  Transform the labels to a one-hot representation.

• properties:

  • depth:

    • type: integer
    • default: None
    • minimum: 1

  • always_apply:

    • type: boolean
    • default: False

  • p:

    • type: number
    • default: 1.0
    • minimum: 0
    • maximum: 1

OneOf

• type: object

• category: Compose

• description:

  Select one of the transforms to apply. Probabilities will be weights.

• properties:

  • transforms:

    • type: array
    • items:
      • type: object

  • p:

    • type: number
    • default: 1.0
    • minimum: 0
    • maximum: 1

RandAugment

• type: object

• category: Pixel level

• description:

  This transform implements the data augmentation policy from this paper: https://arxiv.org/abs/1909.13719

  In summary, there are 12 available augmentations: [Identity, Equalize, Rotate, Solarize, Posterize, RandomContrast, RandomBrightness, GaussianBlur, ShearX, ShearY, TranslateX, TranslateY]. Out of this list, every time that the transform is applied, a set of number_transforms is chosen, e.g. [Posterize, TranslateY], and these transforms are applied with magnitude magnitude.

Example

-------
>>> transform = tr.RandAugment(number_transforms=2, magnitude=13)>>> transformed = transform(image=image, bboxes=bboxes, bboxes_labels=bboxes_labels, mask=mask)

• properties:

  • number_transforms:

    • type: integer
    • default: 2
    • minimum: 1
    • maximum: 12

  • magnitude:

    • type: number
    • default: 4
    • minimum: 0
    • maximum: 10

  • always_apply:

    • type: boolean
    • default: False

  • p:

    • type: number
    • default: 1.0
    • minimum: 0
    • maximum: 1

RandomBrightness

• type: object

• category: Pixel level

• description:

  Randomly change brightness of the input image.

• properties:

  • lower_limit:

    • type: number
    • default: -0.2
    • minimum: -1
    • maximum: 1

  • upper_limit:

    • type: number
    • default: 0.2
    • minimum: -1
    • maximum: 1

  • always_apply:

    • type: boolean
    • default: False

  • p:

    • type: number
    • default: 0.5
    • minimum: 0
    • maximum: 1

RandomBrightnessContrast

• type: object

• category: Pixel level

• description:

  Randomly change brightness and contrast of the input image.

• properties:

  • brightness_lower_limit:

    • type: number
    • default: -0.2
    • minimum: -1
    • maximum: 1

  • brightness_upper_limit:

    • type: number
    • default: 0.2
    • minimum: -1
    • maximum: 1

  • contrast_lower_limit:

    • type: number
    • default: -0.2
    • minimum: -1
    • maximum: 1

  • contrast_upper_limit:

    • type: number
    • default: 0.2
    • minimum: -1
    • maximum: 1

  • always_apply:

    • type: boolean
    • default: False

  • p:

    • type: number
    • default: 0.5
    • minimum: 0
    • maximum: 1

RandomContrast

• type: object

• category: Pixel level

• description:

  Randomly change contrast of the input image.

• properties:

  • lower_limit:

    • type: number
    • default: -0.2
    • minimum: -1
    • maximum: 1

  • upper_limit:

    • type: number
    • default: 0.2
    • minimum: -1
    • maximum: 1

  • always_apply:

    • type: boolean
    • default: False

  • p:

    • type: number
    • default: 0.5
    • minimum: 0
    • maximum: 1

RandomCrop

• type: object

• category: Spatial level

• description:

  Crop a random rectangle of the input.

• properties:

  • height:

    • type: integer
    • default: None
    • minimum: 0

  • width:

    • type: integer
    • default: None
    • minimum: 0

  • always_apply:

    • type: boolean
    • default: False

  • p:

    • type: number
    • default: 1.0
    • minimum: 0
    • maximum: 1

RandomRotate90

• type: object

• category: Spatial level

• description:

  Randomly rotate the input by 90 degrees zero or more times.

• properties:

  • p:
    • type: number
    • default: 0.5
    • minimum: 0
    • maximum: 1

RandomScale

• type: object

• category: Spatial level

• description:

  Randomly resize the input. Output image size is different from the input size.

• properties:

  • lower_limit:

    • type: number
    • default: 0.9
    • minimum: 0
    • maximum: 2

  • upper_limit:

    • type: number
    • default: 1.1
    • minimum: 0
    • maximum: 4

  • interpolation:

    • type: string
    • default: LINEAR
    • enum: ['LINEAR', 'AREA', 'CUBIC', 'NEAREST', 'LANCZOS4']

  • always_apply:

    • type: boolean
    • default: False

  • p:

    • type: number
    • default: 0.5
    • minimum: 0
    • maximum: 1

RemapLabelValues

• type: object

• category: Labels only

• description:

  Transforms the values in the labels given a mapping between current values and desired values.

• properties:

  • label_map:

    • type: object

  • always_apply:

    • type: boolean
    • default: False

  • p:

    • type: number
    • default: 1.0
    • minimum: 0
    • maximum: 1

Resize

• type: object

• category: Spatial level

• description:

  Resize the input to the given height and width.

• properties:

  • height:

    • type: integer
    • default: None
    • minimum: 1

  • width:

    • type: integer
    • default: None
    • minimum: 1

  • always_apply:

    • type: boolean
    • default: False

  • p:

    • type: number
    • default: 1.0
    • minimum: 0
    • maximum: 1

Rotate

• type: object

• category: Spatial level

• description:

  Rotate the input by an angle selected randomly from the uniform distribution.

• properties:

  • lower_limit:

    • type: integer
    • default: -90
    • minimum: -180
    • maximum: 0

  • upper_limit:

    • type: integer
    • default: 90
    • minimum: 0
    • maximum: 180

  • interpolation:

    • type: string
    • default: LINEAR
    • enum: ['LINEAR', 'AREA', 'CUBIC', 'NEAREST', 'LANCZOS4']

  • border_mode:

    • type: string
    • default: REFLECT_101
    • enum: ['CONSTANT', 'REPLICATE', 'REFLECT', 'WRAP', 'REFLECT_101']

  • always_apply:

    • type: boolean
    • default: False

  • p:

    • type: number
    • default: 0.5
    • minimum: 0
    • maximum: 1

ShiftScaleRotate

• type: object

• category: Spatial level

• description:

  Randomly apply affine transforms: translate, scale and rotate the input.

• properties:

  • shift_lower_limit:

    • type: number
    • default: -0.0625
    • minimum: -1
    • maximum: 0

  • shift_upper_limit:

    • type: number
    • default: 0.0625
    • minimum: 0
    • maximum: 1

  • scale_lower_limit:

    • type: number
    • default: -0.1
    • maximum: 0

  • scale_upper_limit:

    • type: number
    • default: 0.1
    • minimum: 0

  • rotate_lower_limit:

    • type: integer
    • default: -45
    • minimum: -360
    • maximum: 0

  • rotate_upper_limit:

    • type: integer
    • default: 45
    • minimum: 0
    • maximum: 360

  • interpolation:

    • type: string
    • default: LINEAR
    • enum: ['LINEAR', 'AREA', 'CUBIC', 'NEAREST', 'LANCZOS4']

  • border_mode:

    • type: string
    • default: CONSTANT
    • enum: ['CONSTANT', 'REPLICATE', 'REFLECT', 'WRAP', 'REFLECT_101']

  • always_apply:

    • type: boolean
    • default: False

  • p:

    • type: number
    • default: 0.5
    • minimum: 0
    • maximum: 1

VerticalFlip

• type: object

• category: Spatial level

• description:

  Flips the input horizontally.

• properties:

  • p:
    • type: number
    • default: 0.5
    • minimum: 0
    • maximum: 1

Available Post-processing Transforms

BinaryPixelThresholdRule

• type: object

• category: Postprocessing

• description:

  Converts a binary segmentation mask to image level OK/NG result based on percentage of pixels greater than a probability threshold

• properties:

  • probability_threshold:

    • type: number
    • default: None
    • minimum: 0
    • maximum: 1

  • area_percent_threshold:

    • type: number
    • default: None
    • minimum: 0
    • maximum: 1

  • always_apply:

    • type: boolean
    • default: False

  • p:

    • type: number
    • default: 1.0
    • minimum: 0
    • maximum: 1

FilterBBoxesDimensions

• type: object

• category: Postprocessing

• description:

  Transform that filters boxes of a particular class depending on the dimensions.

• properties:

  • defect_map:

    • type: string

  • height_threshold:

    • type: integer
    • default: None
    • minimum: 0

  • width_threshold:

    • type: integer
    • default: None
    • minimum: 0

  • to_ignore:

    • type: array
    • items:
      • type: string

  • filter_slim:

    • type: boolean
    • default: True

  • filter_short:

    • type: boolean
    • default: True

  • always_apply:

    • type: boolean
    • default: False

  • p:

    • type: number
    • default: 1.0
    • minimum: 0
    • maximum: 1

FilterMaskDimensions

• type: object

• category: Postprocessing

• description:

  Zeroes out mask values above 0 whose area is below a certain threshold. The threshold can be applied per category.

• properties:

  • area_percent_threshold:

    • anyOf:

      • type: number

        • default: 0.0
        • minimum: 0
        • maximum: 1

      • type: array

        • minItems: 1
        • default: [0.0]
        • items:
          • type: number
          • minimum: 0

  • always_apply:

    • type: boolean
    • default: False

  • p:

    • type: number
    • default: 1.0
    • minimum: 0
    • maximum: 1

FilterMaskScore

• type: object

• category: Postprocessing

• description:

  Zeroes out mask_scores that are below a certain threshold. The threshold can be applied per category.

• properties:

  • score_threshold:

    • anyOf:

      • type: number

        • default: 0.5
        • minimum: 0
        • maximum: 1

      • type: array

        • minItems: 1
        • default: [0.5]
        • items:
          • type: number
          • minimum: 0

  • always_apply:

    • type: boolean
    • default: False

  • p:

    • type: number
    • default: 1.0
    • minimum: 0
    • maximum: 1

ObjectDetectionToClassification

• type: object

• category: Postprocessing

• description:

  Transforms an object detection output into a classification output. by simply taking the bounding box with the highest score. If the key 'label' is passed keyword argument, this transform will ignore it.

• properties:

  • always_apply:

    • type: boolean
    • default: False

  • p:

    • type: number
    • default: 1.0
    • minimum: 0
    • maximum: 1

SegmentationToClassification

• type: object

• category: Postprocessing

• description:

  Transforms a multiclass segmentation output into a classification output.

  The segmentation output can be in terms of both probabilities (using the 'mask_scores' keyword argument) and indices (using 'mask' keyword argument). If both 'mask' and 'mask_scores' are provided, we will going to use 'mask' to compute this transform. If the key 'label' is passed keyword argument, this transform will ignore it.

  Beware that this transform doesn't check that the pixels are connected before computing the predicted area.

• properties:

  • always_apply:

    • type: boolean
    • default: False

  • p:

    • type: number
    • default: 1.0
    • minimum: 0
    • maximum: 1

WarpPerspective

• type: object

• category: Postprocessing

• description:

  WarpPerspective performs 4 point transformation of a region in the input image. Reference - https://www.pyimagesearch.com/2014/08/25/4-point-opencv-getperspective-transform-example/

• properties:

  • original_roi:

    • type: array
    • default: 0
    • items:
      • type: number
      • minimum: 0
      • maximum: 1

  • dest_roi:

    • type: array
    • default: 0
    • items:
      • type: number
      • minimum: 0
      • maximum: 1

  • always_apply:

    • type: boolean
    • default: False

  • p:

    • type: number
    • default: 1.0
    • minimum: 0
    • maximum: 1

Custom Transform

CustomTransform

• type: object

• category: Others

• description:

  Custom Transform takes any callable class.

• properties:

  • transform:

    • type: string

  • params:

    • type: object

  • always_apply:

    • type: boolean
    • default: False

  • p:

    • type: number
    • default: 1.0
    • minimum: 0
    • maximum: 1