Video Content Moderation Built with Amazon Rekognition - Ensuring the Safety of UGC Platforms
Automatically moderates videos on UGC platforms and automates media workflows with face search and segment detection. Introduces an event-driven pipeline with S3 / Lambda.
Overview of the Rekognition Video API
Rekognition's Video API analyzes video files stored in S3 asynchronously. Analysis is started with Start APIs such as StartContentModeration, StartLabelDetection, and StartFaceSearch, and a completion notification is sent to an SNS topic. After completion, the results are retrieved with the corresponding Get API. Because the results include a timestamp for each frame, you can identify what was detected at which point in the video. The supported video formats are MOV and MP4 encoded with H.264, with a maximum file size of 10 GB and a maximum video length of 6 hours.
Implementing Content Moderation
On UGC (user-generated content) platforms, videos uploaded by users need to be checked automatically for inappropriate content. A pipeline is built in which an upload of a video to S3 triggers Lambda to call StartContentModeration, the completion is detected through an SNS notification, and the results are evaluated. The detection results include the category of the inappropriate content and a confidence score. A confidence threshold is set, and a workflow is designed that either automatically makes videos with detections at or above the threshold private or sends them to a human review queue.
Face Search and Segment Detection
If you register face images in a face collection in advance, you can use StartFaceSearch to detect the scenes in which registered faces appear in a video. This can be used for searching the scenes in which performers appear at media companies and for tracking specific individuals in security use cases. Segment detection (StartSegmentDetection) is a feature that analyzes the structure of a video, and it automatically identifies shot boundaries (changes of camera cuts), black frames, color bars, slates, studio logos, and end credits. It is effective for automating video editing and detecting ad insertion points at broadcasters and streaming services.
Rekognition Video Pricing
Pricing for the Rekognition Video API is based on the number of minutes of video processed. Content moderation, label detection, and face search are also billed by the minutes of video processed. When processing large volumes of video on a UGC platform, you can keep costs down through optimizations such as setting confidence thresholds appropriately to reduce unnecessary analysis and judging short videos by their thumbnail images with the Image API.
Real-Time Analysis of Live Streams
In addition to asynchronous analysis of stored files, there is also a path for analyzing streaming video such as live streams in real time. In this configuration, video ingested into Kinesis Video Streams is analyzed by a Rekognition stream processor. The available analyses are face search and label detection (the streaming events feature is not available to new customers from April 30, 2026, and existing users can continue to use it). Because content moderation cannot be used with stream processors, judgment of video during a stream is designed within the scope of face search and label detection. Batch processing, which analyzes recordings together after the fact, and real-time processing, which judges during the stream, differ in nature, so they are used selectively according to requirements. Whether immediacy is required is the dividing line for the choice.
Best Practices for Moderation Design
The basic approach to content moderation is to combine it with human review rather than leaving it entirely to machines. Because the detection categories are hierarchical, you decide at which level of granularity to judge according to the use case. A two-stage approach is practical: set thresholds on the confidence score, automatically block content that is clearly problematic, and send borderline content to a human review queue. This human review flow is prepared in advance as a mechanism. It is also important to record the basis and results of judgments so that they can be verified later.
Privacy and Responsible Use
Features that handle face detection and search require consideration for privacy. Capturing / using footage presupposes obtaining the consent of the people involved and checking applicable laws / guidelines. For collected face data and collections, the retention period and purpose of use are made clear, and they are deleted once they are no longer needed. Because detection results can contain errors, the design does not leave important decisions solely to machine output and includes human confirmation. It is also necessary to note that regulations differ by region. Along with the convenience of the technology, an operational design that protects the rights of users is required.
Choosing Between the Image API and Video, and Costs
Because video analysis is billed according to the number of minutes processed, you assess whether everything needs to be processed as video. For short clips, or when checking representative frames is sufficient, extracting thumbnail images from the video and judging them with the Image API can sometimes be cheaper. A staged design is also effective, in which a lightweight check is performed first at upload time and only suspicious content is sent on to detailed video analysis. By narrowing the analyzed categories to the necessary ones and keeping the length of the processed videos down, you can manage costs even on platforms that handle large volumes of content.
Summary
The Rekognition Video API is a service that achieves automatic analysis of video content in a serverless way. Content moderation helps improve the safety of UGC platforms, and you can automate media workflows with face search and segment detection. With an event-driven architecture that combines S3 and Lambda, you can automate the process from video upload to analysis completion.
References (Official AWS Resources)
The primary sources for this page are the official AWS website and documentation. Check the official pages below for the latest specifications and pricing.
- Amazon Rekognition official page
- Amazon S3 official page
- Amazon Kinesis official page
- AWS Documentation (official)
If this page and the official documentation disagree, treat the official documentation as authoritative.