Rekognition
Amazon Rekognition is a computer vision service that analyses images and video through pre-trained deep learning models. It identifies and labels the content of pictures and video without requiring the caller to train a model.
Capabilities
Section titled “Capabilities”- Label detection — objects, scenes and activities present in an image or video
- Content moderation — flagging unsafe or unwanted imagery against a hierarchy of moderation categories
- Face detection and analysis — locating faces and describing attributes such as pose and image quality
- Face comparison and search — matching a face against a collection of previously indexed faces
- Celebrity recognition — identifying well-known people
- Text in image — extracting text overlaid on or photographed within an image
- Streaming video event detection — processing a Kinesis Video Stream and emitting events when configured conditions are met
Image operations are synchronous: a request returns labels and confidence scores directly. Stored-video operations are asynchronous — a job is started, and completion is signalled through Amazon SNS. Streaming video analysis runs continuously against a Kinesis Video Stream.
Practical notes
Section titled “Practical notes”- Every result carries a confidence score. Pick a threshold deliberately for the use case; the right threshold for content moderation is not the right threshold for a photo-tagging feature.
- Face comparison is a similarity measure, not an identity assertion. Treat it as one input to a decision, not the decision.
- Applications that identify people carry data protection obligations. Under UK and EU data protection law, biometric data used to uniquely identify someone is special category data and needs a lawful basis, a DPIA and appropriate safeguards before any deployment.