Improving Accuracy

Overview

Face recognition accuracy depends on several factors — lighting, camera quality, face angle, and matching thresholds. This guide covers practical techniques to maximize recognition reliability.

Registration Best Practices

Capture Multiple Templates

Store 3-5 face templates per user captured under different conditions:

Lighting — One in bright light, one in dim light
Angle — Straight on, slight left, slight right
Distance — Close up and at normal distance

FaceSmash's multi-template matching automatically uses the best-matching template, so variety improves reliability.

Enforce Quality Minimums

Reject low-quality captures during registration. A bad registration template leads to all future logins failing.

// Recommended SDK config for stricter registration quality
const client = createFaceSmash({
  minQualityScore: 0.5,          // Higher than default (0.2)
  minDetectionConfidence: 0.4,   // Higher than default (0.3)
  matchThreshold: 0.45,
});

Guide the User

Show real-time quality feedback during registration:

"Move closer" — Face too small
"Face the camera" — Head angle too large
"Find better lighting" — Lighting score low
"Hold still" — Blur detected

Adaptive Thresholds

FaceSmash automatically adjusts match thresholds based on:

Factor	Effect
Poor lighting	Threshold lowered by 0.05 (more lenient)
Good lighting	Threshold raised by 0.02 (slightly stricter)
Many successful logins	Confidence boost applied
High success rate	Model trusts the match more

Understanding Similarity Scores

1.0  ┃ Perfect match (same image)
     ┃
0.8  ┃ Very strong match
     ┃
0.6  ┃ Strong match — recommended for high-security
     ┃
0.45 ┃ ← Default threshold (balanced)
     ┃
0.35 ┃ ← Minimum threshold (poor lighting fallback)
     ┃
0.2  ┃ Weak — likely different people
     ┃
0.0  ┃ No similarity

If a legitimate user fails to login, check these in order:

Lighting — Is the room well-lit? Avoid backlighting (window behind user)
Camera — Is the webcam clean and focused?
Face angle — Is the user facing the camera directly?
Glasses/mask — Significant face occlusion reduces accuracy
Template quality — Were registration templates captured in good conditions?

Environment Tips

Lighting

Condition	Impact	Solution
Backlit (window behind)	Very bad — face appears dark	Face the light source
Overhead fluorescent	Moderate — harsh shadows	Add a desk lamp
Natural daylight	Best — even illumination	Ideal for registration
Very dim	Bad — low contrast	Turn on more lights

Camera

Built-in webcam — Usually sufficient, 720p or higher
External webcam — Often better quality and angle
Phone camera — Excellent quality, good for registration

Distance

The ideal face-to-camera distance is 40-80cm (roughly arm's length). The face should fill about 25-35% of the frame.

Technical Tuning

Threshold Selection Guide

Use Case	Match Threshold	Duplicate Threshold
Consumer app (convenience)	0.40 - 0.45	0.75
Business app (balanced)	0.45 - 0.55	0.80
High security (strict)	0.55 - 0.65	0.85

Descriptor Quality

The 128-dimensional face descriptor is the core of matching. Quality depends on:

Detection model — SSD MobileNet v1 (default) is more accurate than TinyFaceDetector
Input resolution — Higher resolution captures more facial detail
Face alignment — Landmarks-based alignment improves descriptor quality
Multiple captures — Averaging descriptors from multiple frames reduces noise

Multi-Template Matching

When a user has multiple templates, FaceSmash compares against all of them and uses the best match:

Template 1 (bright lighting):  similarity = 0.52
Template 2 (dim lighting):     similarity = 0.67  ← best
Template 3 (slight angle):     similarity = 0.48

Result: similarity = 0.67, isMatch = true (threshold 0.45)

This is why storing templates from different conditions dramatically improves reliability.

On this page