Skip to content

Crop Requirements

Guidelines for creating properly sized face crops for the fast-check-crops endpoint.

In plain terms

When using the crops endpoint, each face image must be 224×224 pixels and include context around the face—hair, neck, some background. The face should fill only ~30% of the crop. Too-tight crops (face fills >50%) reduce accuracy; the API may return a warning.

Overview

The fast-check-crops endpoint requires pre-cropped face images that include significant context around the face. This allows for faster processing since the server skips face detection.

Cropping Process

Original
Frame
→
Face
Detected
→
3x Face
Region
→
224×224
PNG
  1. Detect the face bounding box in the original frame
  2. Calculate the center of the face bounding box
  3. Expand to 3× the face size (use the larger of width/height)
  4. Create a square crop centered on the face
  5. Resize the crop to exactly 224×224 pixels
  6. Export as PNG (base64 encoded)

Critical: Include context around the face

The crop must include significant context around the face (hair, neck, some background). If the face fills more than 50% of the crop, the model's accuracy degrades significantly.

Correct vs Incorrect Crops

❌ Too Tight (Wrong)

  • Face fills >50% of crop
  • Little or no context visible
  • May trigger warning in response
  • Accuracy significantly degraded

✅ Correct

  • Face is ~30% of crop
  • Hair, neck, and background visible
  • Optimal context for detection
  • Maximum accuracy achieved

Specifications

Specification Value
Output Size 224 × 224 pixels (square)
Format PNG, base64 encoded
Face Size ~30% of crop area
Expansion Factor 3× the detected face size
Platform Library Description
JavaScript @mediapipe/tasks-vision Browser-based face detection
Python mediapipe Google's MediaPipe face detection

See the JavaScript Examples and Python Examples for complete crop implementations.