Amazon SageMaker Ground Truth is an AWS data-labeling service used to build high-quality labeled datasets for machine-learning workflows. It helps teams label images, text, videos, and 3D point-cloud data so the resulting annotations can be used to train and evaluate machine-learning models.
What problem does Ground Truth solve?
Machine-learning models learn from examples. In supervised learning, those examples need labels that describe the correct class, object, text category, or other target. Ground Truth provides workflows for assigning and reviewing those labels at scale.
For example, an image-recognition team may need thousands of images annotated with bounding boxes around vehicles. A text-classification team may need documents categorized by topic or sentiment. Ground Truth helps organize the labeling work and produces labeled output that can be used in downstream ML workflows.
How a labeling workflow works
- Prepare input data: store the source objects, commonly in Amazon S3, and create a manifest that identifies the data to label.
- Choose the task type: select a supported built-in task such as image classification, object detection, semantic segmentation, text classification, named entity recognition, video labeling, or 3D point-cloud labeling.
- Configure workers and instructions: define the workforce, worker task template, labeling instructions, and access permissions.
- Create a labeling job: Ground Truth distributes tasks and collects annotations. Some supported task types can use automated data labeling to reduce manual effort.
- Review and use the output: inspect labels and use the output manifest and annotations in the model-training pipeline.
Workforce options
Ground Truth supports different ways to source human annotators, including a private workforce, third-party vendors, and (where supported and available) Amazon Mechanical Turk. Ground Truth Plus is a managed option in which AWS helps create and manage data-labeling workflows and the workforce. Availability and eligibility can vary, so verify current AWS service guidance.
Types of data-labeling tasks
| Task type | What workers do | Example |
|---|---|---|
| Image classification | Assign one or more labels to an image. | Classify a photograph as containing a car, truck, or pedestrian. |
| Object detection | Draw bounding boxes around objects. | Mark every vehicle in a street image. |
| Semantic segmentation | Label pixels or regions by class. | Separate road, building, and vegetation pixels. |
| Text classification | Assign categories to text. | Classify support messages by issue type. |
| Named entity recognition | Identify entities in text. | Mark organization names, dates, or locations. |
| Video and 3D point clouds | Label frames, track objects, or annotate 3D data. | Track vehicles across video frames or label objects in LiDAR data. |
Ground Truth vs. A2I vs. Clarify
| Service | Main purpose | Exam clue |
|---|---|---|
| SageMaker Ground Truth | Label datasets to prepare training data. | Workers annotate images, text, video, or 3D data for ML training. |
| Amazon Augmented AI (A2I) | Human review of selected ML predictions or outputs in supported workflows. | Send low-confidence or selected inference results to a human reviewer. |
| SageMaker Clarify | Help identify potential bias and explain model predictions. | Analyze bias metrics or feature attribution. |
Memory hook: Ground Truth labels data before or as part of model development; A2I routes selected results for human review; Clarify focuses on bias and explainability. These services solve different problems.
Key concepts and exam traps
- Ground Truth is for labeling, not model hosting. It creates annotations and labeled datasets; it does not replace the service used to train or deploy a model.
- Input and output data: input data is supplied for labeling, often through Amazon S3 and a manifest. Output includes labels and metadata in an output manifest.
- Human and automated labeling: some supported task types can use automated data labeling, but the workflow still needs appropriate quality controls.
- Ground Truth Plus: this is the managed labeling option; do not confuse it with a model-training algorithm.
- Choose A2I for inference review: when the requirement is to send selected predictions to people, distinguish that from creating a labeled training dataset.
- Check current availability: Ground Truth is no longer open to new customers according to AWS documentation. Do not present it as an unrestricted new-customer option.
Test yourself
Question: A machine-learning team has thousands of unlabeled images and needs workers to draw bounding boxes around objects so the annotations can be used to train an object-detection model. Which AWS service is the best match?
- Amazon SageMaker Clarify
- Amazon Augmented AI (A2I)
- Amazon SageMaker Ground Truth
- Amazon CloudWatch
Show answer and explanation
Correct answer: C. Ground Truth provides data-labeling workflows, including object detection annotations such as bounding boxes. Clarify focuses on bias and explainability; A2I is for human review of selected predictions; CloudWatch monitors resources and applications.
Related AWS topics
Amazon Augmented AI (A2I) · Amazon SageMaker Clarify · Amazon SageMaker Model Cards · Amazon SageMaker JumpStart
Service availability and features can change. Use current AWS documentation when making implementation decisions.