Image Level Augmentations

The Data Generation module allows developers to introduce augmentations to the collected images. That is, enriching base dataset with operations like vertical and horizontal flipping. This allows for more variety in the collected data and leads to a more rounded dataset at the end. These augmentations can also be used to induce edge-case scenarios to the dataset; again for obtaining datasets that more closely represents the real-world scenarios.

The augmentation module can be enabled using the following param:

"augmentation-spec": {
        "2DBBox": {
            "enabled": true
        },
}

Currently, the following properties can be augmented through the module:

An example for configuring these augmentations at an asset level can be found in datacollector_config.jsonc

Refer to augmentation APIs section from api.md for the API reference for the augmentation module.

Horizontal Flipping

Allows for the flipping of the input horizontally around the y-axis. This feature can be configured using the following param:

"horizontal-flip": { //Flip horizontally around the y-axis.
    "enabled": true,
    "p": 1 //probability of applying the transform. Default: 1.
}

Vertical Flipping

Allows for the flipping of the input horizontally around the y-axis. This feature can be configured using the following param:

"vertical-flip": { //Flip horizontally around the y-axis.
    "enabled": true,
    "p": 1 //probability of applying the transform. Default: 1.
}

Random Cropping

Crops a random part of the input. This feature can be configured using the following param:

"crop": { //Crop a random part of the input
    "enabled": true,
    "crop-height": 224, //Height of cropped image
    "crop-width": 224, //Width of cropped image
    "p": 1 //probability of applying the transform. Default: 1.
},

Random Rotation

Rotate the input by an angle selected randomly from the uniform distribution. This feature can be configured using the following param:

"rotate": { //Rotate the input by an angle selected randomly from the uniform distribution
    "enabled": true,
    "angular-limit": 180, //Rotation angle randomly selected from range(-limit, limit)
    "p": 1 //probability of applying the transform. Default: 1.
},

Affine Transformation

Apply an affine transformation to the input image. This feature can be configured using the following param:

"affine-transform": { //Apply affine transformations to image
    "enabled": true,
    "translate-percent": 20, //Translation as a fraction of the image height/width
    "scale": 0.5, //Scaling factor in. Param Range - (0,1)
    "shear": 30, //Shear image along x-axis with angle in range(-angle, angle)
    "p": 1
},

Brightness-Contrast

Randomly change brightness and contrast of the input image. This feature can be configured using the following param:

"brightness-contrast": { //Randomly change brightness and contrast of the input image
    "enabled": true,
    "brightness-limit": 0.5, //factor range for changing brightness. Select value from (0,1)
    "contrast-limit": 0.2, //factor range for changing contrast. Select value from (0,1)
    "p": 1
},

Hue-Saturation-Value

Randomly change hue, saturation and value of the input image. This feature can be configured using the following param:

"hue-saturation-value": { //Randomly change hue, saturation and value of the input image
    "enabled": true,
    "hue-shift-limit": 20, //hue shift randomly selected from range(-limit, limit)
    "saturation-shift-limit": 50, //sat shift randomly selected from range(-limit, limit)
    "value-shift-limit": 30, //value shift randomly selected from range(-limit, limit)
    "p": 1
},

Motion Blur

Apply motion blur to the input image using a random-sized kernel. This feature can be configured using the following param:

"motion-blur": { //Apply motion blur to the input image using a random-sized kernel
    "enabled": true,
    "blur-limit": 21, //maximum kernel size for blurring the input image. Must be a odd number
    "p": 1
},

Gaussian Noise

Apply gaussian noise to the input image. This feature can be configured using the following param:

"gaussian-noise": { //Apply gaussian noise to the input image
    "enabled": true,
    "variance-limit": 200, //variance range for noise 
    "p": 1
}