AWS Rekognition
Computer vision API that detects objects, scenes, faces, text, and explicit content in images and videos, and supports facial comparison and search against collections.
Score Breakdown
⚙ Agent Friendliness
🔒 Security
Images submitted inline are not stored by Rekognition. S3 input images remain under customer control. Face collection vectors are stored encrypted in AWS-managed storage. HIPAA-eligible. VPC endpoints available. Biometric data handling requires careful legal review regardless of AWS security posture.
⚡ Reliability
Best When
You need reliable, off-the-shelf computer vision for common categories (objects, scenes, explicit content, text in images) without training your own model.
Avoid When
Your use case involves facial recognition of private individuals or you operate in jurisdictions with facial recognition restrictions — the legal and ethical risk outweighs the technical convenience.
Use Cases
- • Agents performing content moderation on user-uploaded images — detecting explicit, violent, or policy-violating content
- • Scene and object detection for agents that need to understand image context before processing
- • Extracting text from images (signs, labels, screenshots) as a complement to document processing
- • Video analysis workflows — detecting scenes, people, or activities in recorded video with timestamp metadata
Not For
- • Production facial recognition without careful legal review — several US states and jurisdictions restrict automated facial recognition use
- • Real-time video analysis at scale without provisioned infrastructure — streaming video analysis costs accumulate quickly
- • Fine-grained product recognition or specialized domain objects — custom labels require a training dataset and additional cost
Interface
Authentication
AWS SigV4 signing via IAM credentials or roles. Face-related operations (IndexFaces, SearchFacesByImage, CompareFaces) require separate IAM policy acknowledgment due to biometric data sensitivity. Rekognition Custom Labels requires additional permissions for model training.
Pricing
Costs depend heavily on which features are used — content moderation is cheaper than custom labels. Video analysis pricing is per minute of video, not per detection. Streaming video analysis requires a dedicated inference unit.
Agent Metadata
Known Gotchas
- ⚠ Face collection operations require explicit IAM policy acknowledgment ('rekognition:SearchFacesByImage' plus a separate consent policy) — missing this causes confusing AccessDeniedException errors
- ⚠ DetectText (text in images) is optimized for signs and printed text, not document OCR — use Textract for documents
- ⚠ Image size limit is 5MB for inline bytes, 15MB for S3 references — agents must check before submitting
- ⚠ Confidence thresholds for content moderation must be tuned per use case — default 50% threshold is too low for most production applications
- ⚠ Stored video analysis is asynchronous with SNS/SQS notification or polling — no synchronous video API exists
Alternatives
Full Evaluation Report
Detailed scoring breakdown, competitive positioning, security analysis, and improvement recommendations for AWS Rekognition.
Scores are editorial opinions as of 2026-03-06.