calesthio/generative-media-skills/amazon-rekognition

Amazon Rekognition for Production Image and Video Understanding

Amazon Rekognition is an AWS computer-vision service for analyzing existing images and videos. Use it to extract metadata, screen user-generated content, build searchable media libraries, support media operations, or add narrowly governed identity-verification steps. Do not use this skill for generating images, editing pixels, creating synthetic media, or judging whether a human reference is ethically usable for generation; those are separate generation and reference-analysis tasks.

This skill treats Rekognition as an analysis system that returns probabilistic metadata. It is not a replacement for editorial review, legal review, consent management, or a human moderation policy.

Evidence Legend

• Documented fact: Stated in AWS documentation or AWS-owned service pages. Consequential facts include source links and volatile facts include a verification date.
• Production heuristic: Operational guidance inferred from documented behavior and common production requirements; validate against the application, jurisdiction, and risk tolerance.
• Empirical observation: A result from a reproducible test. This skill does not include first-party experiments; when you test, record dataset, date, region, API version/model version, thresholds, and confusion matrix.

Activation Boundaries

Use Rekognition when the user asks for:

• Image labels, objects, scenes, concepts, celebrities, image properties, or bounding boxes.
• OCR in images or stored video, especially signs, packaging, screenshots, thumbnails, or metadata extraction.
• Moderation or brand-safety screening for images, stored video, or UGC workflows.
• Custom object, logo, scene, or concept detection with Amazon Rekognition Custom Labels.
• Custom moderation adapters for domain-specific moderation performance.
• Stored-video analysis: labels, moderation, text, celebrities, faces, people tracking, face search, or segment detection.
• Searchable media libraries from Rekognition metadata stored in a search index or database.
• AWS-native async pipelines using S3, IAM, SNS, SQS, Lambda, Kinesis, CloudTrail, and cost controls.
• Face liveness or face search only when the use case is explicit identity verification, access control, fraud prevention, or user-consented account recovery.

Do not activate for:

• Image generation, video generation, inpainting, upscaling, restoration, or prompt writing for generative models.
• Human likeness reference analysis for creative generation, casting, style transfer, or impersonation. Keep that separate from Rekognition's detection/search APIs.
• General OCR pipelines needing document layout, tables, handwriting-heavy forms, or multi-page documents; consider Amazon Textract or a document OCR system.
• Surveillance normalization: broad public-space identification, covert tracking, protest monitoring, or open-ended watchlists should be refused or escalated to policy/legal review. Rekognition can technically search faces, but this skill does not make that an ordinary production pattern.

Capability Map

Documented fact, verified 2026-07-11: AWS describes Rekognition as a cloud image and video analysis service with APIs for object/scene/concept labels, text detection, unsafe-content moderation, celebrity recognition, face analysis/comparison/search, Custom Labels, image properties, stored-video analysis, streaming-video events, and Face Liveness. Source: AWS Rekognition Developer Guide, "What is Amazon Rekognition?" https://docs.aws.amazon.com/rekognition/latest/dg/what-is.html

Image Analysis

• DetectLabels returns general labels for objects, scenes, activities, concepts, parents, aliases, categories, confidence scores, and bounding boxes for common object instances.
• DetectLabels can also request IMAGE_PROPERTIES, returning quality scores such as brightness, sharpness, and contrast plus dominant colors for the image, foreground, background, and detected objects when available.
• DetectText returns detected words and lines, confidence, line/word relationships, and geometry as bounding boxes and polygons.
• DetectModerationLabels returns unsafe-content labels and confidence scores for image moderation.
• DetectProtectiveEquipment can detect PPE for safety-compliance workflows.
• RecognizeCelebrities and GetCelebrityInfo support celebrity recognition, but use only when the product explicitly needs celebrity metadata and has rights/policy clearance.
• DetectFaces, CompareFaces, and collection APIs support face attributes, comparison, indexing, and search. Treat these as biometric workflows, not generic tagging.

Documented fact, verified 2026-07-11: Rekognition image non-storage operations do not persist information discovered about the input image, and no input image bytes are persisted by non-storage operations. Storage-based face operations persist face vectors in collections. Source: AWS "Understanding non-storage and storage API operations" https://docs.aws.amazon.com/rekognition/latest/dg/how-it-works-storage-non-storage.html

Stored Video Analysis

Documented fact, verified 2026-07-11: Rekognition Video analyzes videos stored in S3 asynchronously with Start* operations, publishes completion status to SNS, and results are retrieved with Get* operations. Source: AWS "Calling Amazon Rekognition Video operations" https://docs.aws.amazon.com/rekognition/latest/dg/api-video.html

Stored-video operations include:

• StartLabelDetection / GetLabelDetection for labels, events, concepts, and activities over time.
• StartContentModeration / GetContentModeration for unsafe-content events over time.
• StartTextDetection / GetTextDetection for OCR over time.
• StartCelebrityRecognition / GetCelebrityRecognition for celebrity appearances.
• StartFaceDetection / GetFaceDetection for face metadata over time.
• StartPersonTracking / GetPersonTracking for person pathing.
• StartFaceSearch / GetFaceSearch for searching faces in a stored video against a collection. Use only in consented, narrowly scoped identity workflows.
• StartSegmentDetection / GetSegmentDetection for shot boundaries and technical cues.

Documented fact, verified 2026-07-11: Stored video supports H.264 video in MPEG-4 or MOV containers, up to 10 GB and up to 6 hours, with a default maximum of 20 concurrent stored-video jobs per account. Video Get results are retained for 7 days, and pagination-token TTL is 24 hours. Sources: AWS "Working with stored video analysis operations" https://docs.aws.amazon.com/rekognition/latest/dg/video.html, AWS "Guidelines and quotas" https://docs.aws.amazon.com/rekognition/latest/dg/limits.html, AWS "Calling Amazon Rekognition Video operations" https://docs.aws.amazon.com/rekognition/latest/dg/api-video.html

Streaming Video And Bulk Image Analysis Availability

Documented volatile fact, verified 2026-07-11: AWS states that Streaming Video and Bulk Image Analysis are no longer available to new customers effective April 30, 2026. Accounts that used these features within the previous 12 months continue to have access; other Rekognition features are not impacted. Source: AWS "Amazon Rekognition feature availability changes" https://docs.aws.amazon.com/rekognition/latest/dg/rekognition-availability-changes.html

For new production designs, prefer:

• Stored-video analysis for files in S3.
• Frame extraction from streams into S3/EventBridge/Lambda or Kinesis-based custom pipelines that call image APIs, if near-real-time analysis is needed.
• Existing eligible streaming customers may use stream processors with Kinesis Video Streams and Kinesis Data Streams or SNS/S3 outputs, but must verify account access and current regional support before promising delivery.

Bulk Image Analysis should likewise be treated as conditional for existing eligible accounts. For new customers, design an S3 manifest or queue-driven worker pipeline that calls the image APIs directly with idempotency, retries, and output persistence.

Customization Choices

General Labels vs Custom Labels

Documented fact, verified 2026-07-11: AWS recommends standard label detection for common labels and Rekognition Custom Labels when the application needs objects, logos, scenes, or concepts that general label detection does not find. Custom Labels can classify images or detect object locations; it is not designed for analyzing faces, detecting text, or unsafe-content detection. Source: AWS Rekognition Custom Labels Developer Guide, "What is Amazon Rekognition Custom Labels?" https://docs.aws.amazon.com/rekognition/latest/customlabels-dg/what-is.html

Use general DetectLabels when:

• The object/scene vocabulary is common enough to appear in Rekognition's general taxonomy.
• The user needs broad tagging, search facets, or rough media-library indexing.
• Precision can be tuned with MinConfidence, inclusion/exclusion filters, and human review.

Use Custom Labels when:

• The target is proprietary, domain-specific, visually subtle, or absent from the general taxonomy.
• The deliverable requires a stable business vocabulary, such as SKU families, character IDs, industrial defects, sports-team logos, or shelf fixtures.
• You can assemble labeled positive/negative examples and an evaluation set that represents production conditions.

Custom Labels lifecycle:

1. Define the exact label schema and whether each label is image-level classification or object localization.
2. Build datasets from consented, rights-cleared images; split train/test without near-duplicate leakage.
3. Train a project version and evaluate precision, recall, F1, confidence distributions, and failure clusters.
4. Start the project version only when serving predictions; stop it when idle to control cost.
5. Store predictions and human corrections so retraining improves real failure modes rather than synthetic edge cases.
6. Version labels, thresholds, and downstream business rules together.

Moderation Adapters

Documented fact, verified 2026-07-11: Rekognition Custom Moderation lets users train moderation adapters in projects and pass an adapter to DetectModerationLabels to improve performance for specific moderation tasks. Source: AWS "Enhancing accuracy with Custom Moderation" https://docs.aws.amazon.com/rekognition/latest/dg/moderation-custom-moderation.html

Use a moderation adapter when the base moderation model misses or overflags content in a defined domain, such as product imagery, illustrated content, platform-specific policy categories, or recurring false positives. Do not use adapters to bypass policy. The policy taxonomy still needs human owners, appeal paths, and sampled audits.

Searchable Library Architecture

A production searchable library is not just Rekognition output. It is an ingestion, metadata, review, and retrieval system.

Recommended architecture: