Confidence Tuning

This guide explains how to configure confidence thresholds to balance detection sensitivity with false positive reduction.

Understanding Confidence Scores

Every Anava detection includes a confidence score indicating how certain the AI is about its analysis.

Score Ranges

Range	Label	Meaning
0-40%	Very Low	Uncertain, likely incorrect
40-60%	Low	Some indicators, needs verification
60-75%	Medium	Reasonable confidence
75-90%	High	Likely correct
90-100%	Very High	Almost certain

Where Confidence Appears

1. Object detections - Each object has a confidence score
2. Question answers - Boolean/set answers may include confidence
3. Overall session - Aggregate confidence for the analysis

Threshold Configuration

Profile-Level Threshold

Set in profile settings:

Setting	Description	Default
Confirmation Threshold	Consecutive detections before TTS fires (1-10)	1

When set:

• Requires N consecutive positive detections within a session before TTS triggers
• ONVIF events still emit immediately (not gated)
• Counter resets after TTS fires
• Counter is per-session (cleared when session ends)

Reducing False Positives

Set to 2-3 to require multiple confirmations before voice response. This is especially useful in high-traffic areas where momentary detections may not warrant immediate TTS.

Object-Level Behavior

In skill configuration, each object can have:

• Rapid Eligible - Can trigger immediate response at high confidence
• Trigger Deep Analysis - Low confidence triggers additional analysis

Setting the Right Threshold

By Use Case

Use Case	Recommended	Rationale
Weapon detection	75-85%	Critical but can't miss
Intrusion detection	80-90%	Balance speed and accuracy
PPE compliance	70-80%	Some tolerance acceptable
Queue monitoring	60-70%	Directional data OK
Analytics only	50%+	Capture all data

The Sensitivity Trade-off

Tuning Process

Step 1: Establish Baseline

1. Set threshold to 0 (no filtering)
2. Run for 24-48 hours
3. Review all sessions
4. Calculate baseline metrics

Step 2: Analyze Distribution

Review session confidence scores:

Sessions by Confidence:
├── 90-100%: 45 sessions (all correct)
├── 80-90%: 32 sessions (30 correct, 2 false positive)
├── 70-80%: 28 sessions (20 correct, 8 false positive)
├── 60-70%: 15 sessions (5 correct, 10 false positive)
└── <60%: 8 sessions (1 correct, 7 false positive)

Step 3: Set Initial Threshold

Based on analysis:

• If 80%+ sessions are accurate at 75%, set threshold to 75%
• Adjust based on false positive tolerance

Step 4: Monitor and Adjust

After setting threshold:

1. Monitor for missed detections
2. Review flagged but not alerted sessions
3. Adjust threshold if needed

Advanced Configurations

Multi-Threshold Strategy

Use different thresholds for different responses:

Confidence	Response
90%+	Immediate alarm + recording
75-90%	Recording + email alert
50-75%	Recording only
Under 50%	Log for review

Implementation:

• Create multiple profiles with different thresholds
• Each profile triggers different VMS actions

Rapid Analysis for Critical Detections

For high-priority objects (weapons):

1. Enable Rapid Eligible on the object
2. High-confidence detections skip full analysis
3. Immediate response for clear threats
4. Full analysis continues for context

TTS Threshold

Separate from ONVIF threshold:

Setting	Purpose
Profile Threshold	When to emit ONVIF events
TTS Confidence	When to trigger voice response

Set TTS higher than ONVIF to reduce unnecessary announcements.

Troubleshooting Threshold Issues

Too Many False Positives

Symptoms:

• VMS flooded with events
• Action rules triggering on non-events
• Operator alert fatigue

Solutions:

1. Increase threshold by 5-10%
2. Review false positives for patterns
3. Improve skill prompts
4. Consider pre-filter

Missing Real Events

Symptoms:

• Known events not triggering
• Sessions show detection but no ONVIF
• VMS not receiving alerts

Solutions:

1. Lower threshold by 5-10%
2. Review missed events' confidence scores
3. Improve prompts for clarity
4. Consider separate high-sensitivity profile

Inconsistent Confidence

Symptoms:

• Same scenario gives wildly different scores
• Threshold that worked stops working

Solutions:

1. Improve prompt specificity
2. Add contextual information
3. Check for environmental changes (lighting)
4. Review recent false positives

Confidence by Object Type

Different objects may need different handling:

High-Confidence Objects

Objects where the AI is typically very certain:

• Person (clearly visible human)
• Vehicle (distinct shape)
• Fire/Smoke (obvious visual)

Recommendation: Threshold 80%+

Lower-Confidence Objects

Objects that are harder to classify:

• Weapon (context-dependent)
• PPE items (small, similar items)
• Behaviors (subjective)

Recommendation: Threshold 70-80% with verification

Seasonal and Environmental Adjustments

Lighting Changes

Condition	Adjustment
Winter (low light)	Lower threshold 5-10%
Summer (bright)	Higher threshold possible
Indoor (consistent)	Standard threshold

Weather Impact

Condition	Effect	Adjustment
Rain	Reduced visibility	Lower threshold
Fog	Obscured details	Lower threshold + review
Snow	Reflection issues	Adjust prompts

Best Practices

Start Conservative

Begin with higher threshold and adjust down:

• Easier to lower threshold than raise
• Avoids initial flood of false positives
• Builds operator trust in system

Document Changes

Keep record of:

• Threshold changes and dates
• Reason for change
• Observed impact
• Revert if needed

Regular Review

Schedule threshold review:

• Monthly for active deployments
• After major changes
• Following complaints/issues

Related Topics

• Tuning & Optimization - Overall tuning strategies
• Learning Mode - Active monitoring configuration
• Skills Reference - Object configuration