Model architectures 🏛

Metrics 📊

Solvers / Optimizers 🧮

Scheduler ⏰

Training parameters

Augmentations

DEPLOYMENT

Powered By GitBook

Shift Scale Rotate

Augment your training images by shifting, scaling, and rotating

Parameters

Probability

It is the probability of applying the transformation to the sample images. Higher probability means that the expected number of transforms is higher.

Rotation Range

It defines the range of the rotation of the transformed image.

Scale Factor

The scale factor is used to rescale the contents of the image. The contents of the images are rescaled according to the scale factor.

The scale factor should be between -1 and 1. A negative scale factor zooms out the image and a positive scale factor zooms in the image. The magnitude defines the amount of zooming.

Shift Factor

The shift factor, as the name suggests, is the numeric value used to specify the amount by which the image needs to be shifted both horizontally and vertically.

The shift factor also ranges between -1 and 1. The negative shift factor shifts the pixels leftwards in the horizontal direction and upwards in the vertical direction. The positive shift factor shifts the pixels rightwards in the horizontal direction and downwards in the vertical direction. The magnitude defines the amount of shift.

Advanced options

Extrapolation method

The extrapolation method is used by the tool itself to fill some parts of the image when they are emptied due to shifting or rescaling. When the contents of the image are shrunk, say by a factor of 10, then most of the part of the image is "emptied". To fill that empty space, one can use one of the many extrapolation methods. The available methods are :

Constant

Replicate

Reflect

Wrap

Reflect 101

Augmenting the coin candy with different extrapolation.

The image of the coin candy has been shifted in the range of 0.5 to 0.5 (which means that it has been shifted by 0.5). Scaling and rotation have not been applied.

Note how the shift works. The coin is shifted right with a factor of 0.5 which cuts off half the image and goes down which cuts off another half. Hence, a quarter of the bottom right image is seen.

The constant method replaces the rest of the three quarters with black.

The replicate option fills the space with the nearest pixels.

Wrap, reflect and reflect 101 all result in almost the same type of image. The missing three quarters have been wrapped from the top and left-hand side to create such an image.

Interpolation method

This is another advanced option to choose the interpolation algorithm that is used to resize the image. The available methods are:

Linear Interpolation

Cubic Interpolation

Nearest Pixel

Area

Lanczos algorithm

Padding Value if border mode is constant

If the constant method is selected as the extrapolation method, then the empty space resulted from the transform is filled by the given padding value instead of filling it with black.

Last modified 1mo ago