Project Overview

Goal

Build a serverless backend that automatically detects, matches, and groups faces across uploaded photos using Amazon Rekognition.

Who this is for

• Beginner-friendly: You can follow steps without prior AWS experience (we'll explain concepts + give copy/paste commands).
• Advanced-friendly: We'll also include best practices like least-privilege IAM, idempotency basics, and clean data modeling.

What you'll build

You will build an end-to-end workflow that produces:

• People clusters (each person has a personId)
• Appearances (which photos that person appears in)
• Face thumbnails (cropped faces stored in S3)
• APIs to list people and fetch photos per person (for UI integration)

Example

• People:

  • personId: p_001 → appears in photo_12, photo_13
  • personId: p_002 → appears in photo_14

• Storage:

  • Raw photos → s3://<raw-bucket>/photos-raw/...
  • Crops/thumbnails → s3://<processed-bucket>/people/p_001/...

AWS services used

We will use these AWS services during the build:

• Amazon S3 (raw uploads + thumbnails)
• AWS Lambda (Ingestion + API)
• Amazon Rekognition (DetectFaces, SearchFacesByImage, IndexFaces, Collections)
• Amazon DynamoDB (photos, people, appearances tables)
• Amazon API Gateway (HTTP endpoints for the client)

Architecture

This project uses two Lambdas:

• Ingestion Lambda (event-driven): processes new photos when S3 receives uploads
• API Lambda (client-facing): serves endpoints via API Gateway (list people, get photos for person)

1. A photo is uploaded to S3 (raw bucket)
2. S3 triggers Ingestion Lambda
3. Lambda calls Rekognition
4. Lambda stores metadata in DynamoDB
5. Lambda stores face crops in S3 (processed bucket)
6. Client calls for API Gateway
7. API Lambda reads DynamoDB and returns clusters

What we'll do

We'll build the system in this order:

1. Set up AWS tooling & environment
2. Create S3 buckets + folder structure
3. Create DynamoDB tables (people + appearances)
4. Create a Rekognition Collection
5. Write the Lambda pipeline:
   • detect faces
   • crop faces
   • search collection
   • index new faces
   • group into people
   • write results
6. Package Pillow as a Lambda Layer
7. Secure access using presigned URLs
8. Add logging + troubleshooting + cleanup guide

Common Pitfalls

• Using the wrong AWS region (Rekognition Collection is region-specific)
• Missing IAM permissions for Lambda → Rekognition or Lambda → DynamoDB
• Uploading images with no visible face
• Face crop coordinates confusion (normalized vs pixel values)

We'll address each of these exactly when we reach that step.