Discarding detections?

[jkrauska]

I'm listening live and definately hear these birds.

Why are p89s being discarded?

2024/12/19 13:32:42 New detection: Chestnut-backed Chickadee with confidence: 0.88, source: malgo
2024/12/19 13:32:52 Discarding detection of chestnut-backed chickadee from source malgo as false positive, matched 1/10 times

2024/12/19 13:35:37 New detection: House Finch with confidence: 0.89, source: malgo
2024/12/19 13:35:48 Discarding detection of house finch from source malgo as false positive, matched 1/10 times

[jkrauska]

More details. Is the match count not tracked in a DB?

2024/12/19 13:39:25 New detection: Chestnut-backed Chickadee with confidence: 0.81, source: malgo
2024/12/19 13:39:25 Updating detection: Chestnut-backed Chickadee with confidence: 0.83, source: malgo, count: 2
2024/12/19 13:39:26 Updating detection: Chestnut-backed Chickadee with confidence: 0.87, source: malgo, count: 3
2024/12/19 13:39:35 Discarding detection of chestnut-backed chickadee from source malgo as false positive, matched 7/10 times

2024/12/19 13:39:56 New detection: Chestnut-backed Chickadee with confidence: 0.63, source: malgo
2024/12/19 13:39:56 Updating detection: Chestnut-backed Chickadee with confidence: 0.92, source: malgo, count: 2
2024/12/19 13:40:06 Discarding detection of chestnut-backed chickadee from source malgo as false positive, matched 7/10 times
2024/12/19 13:40:40 New detection: Chestnut-backed Chickadee with confidence: 0.50, source: malgo
2024/12/19 13:40:51 Discarding detection of chestnut-backed chickadee from source malgo as false positive, matched 1/10 times


2024/12/19 13:42:03 New detection: Chestnut-backed Chickadee with confidence: 0.64, source: malgo
2024/12/19 13:42:04 Updating detection: Chestnut-backed Chickadee with confidence: 0.67, source: malgo, count: 2
2024/12/19 13:42:05 Updating detection: Chestnut-backed Chickadee with confidence: 0.71, source: malgo, count: 3
2024/12/19 13:42:14 Discarding detection of chestnut-backed chickadee from source malgo as false positive, matched 3/10 times
2024/12/19 13:42:38 New detection: Chestnut-backed Chickadee with confidence: 0.56, source: malgo
2024/12/19 13:42:48 Discarding detection of chestnut-backed chickadee from source malgo as false positive, matched 2/10 times
2024/12/19 13:42:54 Human detected with confidence 0.056/0.050 from source malgo
2024/12/19 13:43:04 New detection: Orange-crowned Warbler with confidence: 0.82, source: malgo
2024/12/19 13:43:07 New detection: Chestnut-backed Chickadee with confidence: 0.45, source: malgo
2024/12/19 13:43:08 Updating detection: Chestnut-backed Chickadee with confidence: 0.51, source: malgo, count: 2
2024/12/19 13:43:08 Updating detection: Chestnut-backed Chickadee with confidence: 0.63, source: malgo, count: 3
2024/12/19 13:43:15 Discarding detection of orange-crowned warbler from source malgo as false positive, matched 1/10 times
2024/12/19 13:43:18 Discarding detection of chestnut-backed chickadee from source malgo as false positive, matched 3/10 times

[tphakala]

This behavior is controlled by the "deep detection" feature, which tries to eliminate false positives by requiring multiple results (hits) of species by the BirdNET AI model from a 3-second audio segment before accepting detection.

In the logs you provided, the minimum detection count is 10, and the chestnut-backed chickadee was detected only 3 times with confidence exceeding the configured confidence threshold.

The detection count threshold is controlled by the BirdNET overlap value. The code to calculate the minDetections value is in analysis/processor/processor.go:

// pendingDetectionsFlusher runs a goroutine that periodically checks the pending detections
// and flushes them to the worker queue if their deadline has passed.
func (p *Processor) pendingDetectionsFlusher() {
	// Determine minDetections based on Settings.BirdNET.Overlap
	var minDetections int

	// Calculate segment length based on overlap setting, minimum 0.1 seconds
	segmentLength := math.Max(0.1, 3.0-p.Settings.BirdNET.Overlap)
	// Calculate minimum detections needed based on segment length, at least 1
	minDetections = int(math.Max(1, 3/segmentLength))

I have been thinking of adding configuration options that allow more control over the minDetections value, but then again, I also don't want to make things too complicated for users. However, setting BirdNET overlap value to 0 accepts detections on first hit.

[jkrauska]

@tphakala is the "minimum detection count" a global count? I'm guessing that upon daemon restart that these get zero'd out and need to start over?

if so, can we store in sqllite to avoid a slow start?