Preprocessing

The preprocessing section of the configuration file contains the settings that are used to preprocess the images in the dataset, right before passing them to the model.

preprocessing:
  steps:
    - uses: step1
      args:
        arg1: value1
        arg2: value2
        ...
    - ...

The steps list contains all the preprocessing steps that will be applied to the images in the dataset.

Each step has two keys: uses, which contains the name of the step, and args, which contains the arguments that will be passed to the step.

Note

The uses field is case-insensitive, and all hyphens and underscores are ignored, so STEP1, Step1, step-1 and step_1 are all equivalent and will use the same Step1 step.

The transformations are applied in the order they are specified in the configuration file.

Sometimes it is useful to apply a certain preprocessing step only to a specific dataset, for example when the augmentation pipeline has already taken care of some of the preprocessing steps for the training set. If that's the case, the datasets key can be used to specify the datasets that the preprocessing step will be applied to.

preprocessing:
  steps:
    - uses: step1
      datasets: [train, val]
      args:
        arg1: value1
        arg2: value2
        ...
    - ...

The datasets key (if present) is a list of dataset names (train, val, or test) that the preprocessing step will be applied to. If the datasets key is not present, the preprocessing step will be applied to all the datasets.

Preprocessing steps

There are some preprocessing steps already available in Revelio.

Resize

The resize step resizes the images to the specified size.

The resize step has the following arguments:

• width (required int): the width of the resized images.
• height (required int): the height of the resized images.

• algorithm (optional string, default: cubic): the algorithm to use to resize the images. The available algorithms are the same as the ones available in OpenCV:

  • nearest: nearest-neighbor interpolation
  • linear: bilinear interpolation
  • cubic: bicubic interpolation (default)
  • area: resampling using pixel area relation
  • lanczos4: Lanczos interpolation over 8x8 neighborhood

• keep_aspect_ratio (optional bool, default: True): whether to keep the aspect ratio of the images when resizing them.

• fill_mode (optional string, default: constant): the strategy used for filling in newly created pixels, which can appear when using the keep_aspect_ratio option. The available strategies are the same as the ones available in OpenCV:

  • constant: the pixels are filled with black (default, e.g. 000000|abcdefgh|000000)
  • reflect: the pixels are filled with the reflection of the image (e.g. gfedcb|abcdefgh|gfedcba)
  • replicate: the pixels are filled with the last pixel of the image (e.g. aaaaaa|abcdefgh|hhhhhhh)
  • wrap: the pixels are filled with the wrap of the image (e.g. cdefgh|abcdefgh|abcdefg)

To float

The to_float step converts the images to floating point images in the range [0, 1].

The to_float step has no arguments.

Normalize

The normalize step normalizes the images using the specified mean and standard deviation.

The normalize step has the following arguments:

• mean (optional list of floats, default: None): the mean to use to normalize the images, one element per BGR channel.
• std (optional list of floats, default: None): the standard deviation to use to normalize the images, one element per BGR channel.

• preset (optional string, default: None): the preset to use to normalize the images. The available presets are:

  • imagenet: the mean and standard deviation used to normalize the images in the ImageNet dataset.

Either preset or both mean and std must be specified.

Warning

Watch out for the order of the channels in the mean and standard deviation values. The order is BGR (i.e. OpenCV convention), not RGB.