Skip to content

Augmenting Datasets

Augmentation is a powerful way to increase the size, diversity, and balance of a dataset. Innotescus allows users to augment both data and annotations to rebalance datasets and eliminate biases.

Augmentation

Selecting Items for Augmentation

Data can be augmented in two locations. Users can select one or more items on the dataset details page before clicking the augment icon right above it, or users can select one or more points on the project- or task-level datasets tab, right click, and select 'augment X images' in the context menu.

Selecting Dataset Items for Augmentation

Amount of Augmentation

Users must choose the amount of new data they wish to generate with their augmentation step. The defined augmentors will be applied in the designated method to randomly sampled images until the given number of new images has been created.

Adding Augmentors

To add an augmentor to your augmentation operation, simply click 'add augmentor' and click the type of augmentor you'd like to add from the dropdown that appears. Augmentors fall into three categories: color, spatial, and noise/blur.

Crop/Pad

Users can crop or pad along both the X and Y dimensions according to the ranges provided. A padding mode must be selected; if padding is necessary, the 'edge' padding mode will repeat the pixel values at the edge of the image, and the 'constant' mode will give the empty space a value randomly chosen from the range provided.

Flip

Users can apply horizontal and vertical flips to their data by selecting one or both checkboxes when configuring.

Rotate

Once users supply a minimum and maximum angle for the rotation augmentor, a value from that range will be randomly selected and applied to the data. Any blank spaces left due to the rotation will be filled in with a pixel value of 0.

Resize

Users can resize their data along the X and/or Y axes. For this augmentor, users can decide to enter their chosen parameters as pixel values or percentages. Resize will change the size of the image.

Scale

Users can scale their data along the X and/or Y axes. For this augmentor, users can decide to enter their chosen parameters as pixel values or percentages. Scaling an image will not affect its dimensions, only its contents.

Shear

Users can shear their data along the X and/or Y axes, but an amount that is randomly selected from the range of degrees provided.

Translation

Users can translate their data along the X and/or Y axes. For this augmentor, users have the option to enter each range in pixels or percentages.

Brightness Multiplier

Users can multiply the brightness of their data by an amount within the range given. The brightness multiplier must be greater than or equal to 0.

Contrast Multiplier

Users can adjust the contrast of their data by a factor within the range 0.5 to 2.

Hue Multiplier

Users can adjust the hue of their data by a multiplier ranging from -10 to 10.

Additive Gaussian Noise

To add gausian noise to their data, users must provide a range from which the standard deviation for the gaussian distribution will be sampled.

Gaussian Blur

Users can apply a gaussian blur to their data using a gaussian kernel with a standard deviation sampled from the given range.

Motion Blur

Users can apply a motion blur to their data by giving a kernel size for the blurring operation.

Salt and Pepper

Users can replace a percentage of pixels in their data with salt and peper noise. The percentage replaced will be randomly determined from the range provided.

Augmentation Method

Once all augmentors have been defined, users must choose an augmentation method. The below options determine how the defined augmentors are applied:

All Sequentially

This option applies all defined augmentors in their current order to each sampled image. To rearrange the order of augmentors, simply click and drag them around in the list.

All Randomly

This option applies all defined augmentors in random order to each sampled image.

One Randomly

This option chooses one augmentor from the defined list and applies it to one randomly sampled image.

Include Augmented Items

This option determines whether images that have been augmented as part of this operation are eligible to be sampled again and serve as the 'original' for another augmented image. If it is not selected, only the original selection of data can be sampled from during the augmentation operation.

Destination Dataset

Once the entire augmentation scheme has been defined, users must choose the destination for augmented data, which can be the current dataset, an existing dataset, or an entirely new dataset.

Augmenting Annotations

As the final step in the augmentation process, users can decide if they want to maintain and augment existing annotations for the selected data. All annotations left checked will be augmented appropriately.