Choosing the Right Confidence Threshold for YOLO

Trained a YOLOv8s-P2 model for four-leaf clover detection. Got 88.8% mAP@0.5. Now the question: what confidence threshold should the iOS app use?

The trade-off

YOLO outputs a confidence score (0-1) for each detection. The threshold filters what gets shown to users.

Higher threshold (0.5+)	Lower threshold (0.15)
Fewer false positives	More detections
Misses some objects	More false positives
”Only show me sure things"	"Show me anything suspicious”

There’s no free lunch. You’re trading precision for recall.

Finding the sweet spot

Tried running validation at multiple thresholds to plot an F1-confidence curve. On Apple Silicon (MPS), this was painfully slow due to NMS timeout at low thresholds - too many candidate boxes to process.

The faster approach: use the F1 curve from training output. YOLO automatically generates F1_curve.png showing F1 score at each threshold. The peak is your optimal threshold.

If you don’t have that, here’s the general pattern:

Threshold	Use case
0.15-0.20	High recall needed, FP acceptable
0.25	Balanced (YOLO default)
0.40-0.50	High precision, conservative

For my clover app

The use case matters. For clover detection:

• Missing a clover = user disappointment
• False positive = user taps to dismiss

Users can dismiss FPs, but they can’t find missed clovers. So recall matters more.

My pick: 0.25 as default, with a sensitivity slider letting users choose between 0.15 (sensitive) and 0.40 (conservative).

iOS implementation

enum Sensitivity {
    case low       // 0.40 - sure things only
    case medium    // 0.25 - balanced
    case high      // 0.15 - catch everything

    var threshold: Float {
        switch self {
        case .low: return 0.40
        case .medium: return 0.25
        case .high: return 0.15
        }
    }
}

// Filter Vision framework results
func filter(_ observations: [VNRecognizedObjectObservation],
            threshold: Float) -> [VNRecognizedObjectObservation] {
    observations.filter { $0.confidence >= threshold }
}

Takeaway

Start with 0.25. Ship it. Watch user feedback. If “it’s missing clovers” comes up, lower it. If “too many false alarms” - raise it.

The right threshold is whatever makes your users happy, not what maximizes F1 on a validation set.