Augmenting data sets

A deployed model can be improved by using data augmentation to add modified images to the data set. Then, the model can be retrained.

About this task

Data augmentation is the use of filters, such as blur and rotate, to create new versions of existing images or frames. When you use data augmentation, a new data set is created. It contains all of the existing images plus the newly generated images, which are marked as augmented and contain any labels and categories that are identified in the original image. The following considerations apply:
  • Augmentation does not apply to full videos. It can be applied to captured video frames just as it is applied to images.
  • When you use augmentation with auto-label, accept or reject labels before augmenting the data set because auto-label confidence levels are not preserved in augmented images.
  • When you use augmentation with the rotate filter, objects that are labeled with a bounding box might not be tightly bound after rotation. To make sure that the object remains tightly bound after rotation, label it with a polygon instead of a bounding box. For best results, ensure that you use the drawing tools to outline the object correctly.
Note: The maximum number of files that can be uploaded and augmented is 1 million.

Procedure

  1. Open the data set for a deployed model.
  2. Select the images to use for augmentation.
    1. Click Augment data.
      If you select a video, every captured frame is used for augmentation. If you select some frames in a video, only the selected frames are used for augmentation.
  3. Choose any combination of filters to apply to your data set, then click Continue.
    Each filter generates one or more new versions of each selected image. The filters are not cumulative.
    If you select Sharpen and Flip horizontal, six new images are generated: one flipped and five sharpened.
    When you select a filter, you can see an example of what that filter might do to an image. This sample image is not a live preview of the filter. It is an example of what an image might look like with that filter applied. Some filters, such as Blur and Sharpen, have more settings to choose from.
  4. Specify a name for the new data set and click Create data set.
    The new data set, containing the original images, is created immediately. After all processing is completed, the augmented images are added. After the new data set is created, you can train a model based on the new data set.
    Note: For more information about training a model, see Training a model.