Preparing your datasets - Amazon Rekognition
ImagesLabelsTraining and testing datasets

Preparing your datasets

Creating an adapter requires you to provide Rekognition with two datasets, a training dataset and a testing dataset. Each dataset is comprised of provide two elements: images and annotations/labels. The following sections explain what labels and images are used for and how they come together to create datasets.

Images

You will need to train an adapter on representative samples of your images. When you select images for training, try to include at least a few images that demonstrate the expected response for each of the labels you are targeting with your adapter.

To create a training dataset, you need to provide one of the following two image types:

  • Images with False Positive predictions. For example, when a base model predicts that an image has alcohol, but it doesn't.

  • Images with False Negative predictions. For example, when a base model predicts that an image doesn't have alcohol, but it does.

To create a balanced dataset, it is recommended that you provide one of the following two image types:

  • Images with True Positive predictions. For example, when a base model correctly predicts that an image has alcohol. It is recommended to provide these images if you provide False Positive images.

  • Images with True Negative predictions. For example, when a base model correctly predicts that an image doesn't have alcohol. It is recommended to provide these images if you provide False Negative images.

Labels

A label refers to any of the following: objects, events, concepts or activities. For Content Moderation, a label is an instance of content that is inappropriate, unwanted, or offensive.

In the context of creating an adapter by training Rekognition’s base model, when a label is assigned to an image it’s called an annotation. When training an adapter with the Rekognition Console, you’ll use the Console to add annotations to your images by choosing a label and then tagging images that corresponds with the label. Through this process, the model learns to identify elements of your images based on the assigned label. This linking process allows the model to focus on the most relevant content when an adapter is created, leading to improved accuracy for image analysis.

Alternatively, you can provide a manifest files, which contains information on images and the annotations that go with them.

Training and testing datasets

The training dataset is the basis for fine-tuning the model and creating a custom adapter. You must provide an annotated training dataset for the model to learn from. The model learns from this dataset to improve its performance on the type of images you provide.

To improve accuracy, you must create your training dataset by annotation/labeling images. You can accomplish this in two ways:

  • Manual label assignment - You can use the Rekognition Console to create a training dataset by uploading the images you want your dataset to contain and then manually assign labels to these images.

  • Manifest file — You can use a manifest file to train your adapter. The manifest file contains information on the ground-truth annotations for your training and testing images, as well as the location of your training images. You can provide the manifest file when training an adapter using the Rekognition APIs or when using the AWS Console.

The testing dataset is used to evaluate the adapter’s performance after training. To ensure reliable evaluation, the testing dataset is created by using a slice of the original training dataset that the model hasn’t seen before. This process ensures that the adapter’s performance is assessed with new data, creating accurate measurements and metrics. For optimal accuracy improvements see Best practices for training adapters .