How to stop - Shadow Detection

[jeeftor]

Any suggestions on how to reduce these false positives ? I've trained them as "NO" but I'm thinking there should be some other way to reduce them whether it's the percent match - or maybe a movement threshold? I'm not sure.

[inkeep-ai-assistant[bot]]

To reduce false positives from shadow detections, you can adjust several object filter settings in your Frigate configuration:(1)

1. Increase the threshold value

The threshold determines when a tracked object is considered a true positive.(1) Increase this value to require higher confidence scores before an object is confirmed:(1)

objects:
  filters:
    rabbit:
      threshold: 0.85

(1)

2. Increase min_score

The min_score filters out detections below a certain confidence level immediately.(1) Any detection below this value is thrown out as a false positive:(1)

objects:
  filters:
    rabbit:
      min_score: 0.7

(1)

3. Adjust object area filters

Use min_area and max_area to filter detections based on bounding box size.(1) This can help eliminate shadows that are too small or oddly shaped:(1)

objects:
  filters:
    rabbit:
      min_area: 1000
      max_area: 50000

(1)

4. Tune motion detection

Adjust the motion threshold to make motion detection less sensitive to subtle changes like moving shadows:(2)

motion:
  threshold: 35

(2)

Increase the contour_area to ignore smaller motion regions that might be caused by shadows:(2)

motion:
  contour_area: 30

(2)

5. Use motion masks

Create motion masks to exclude areas where shadows regularly appear.(2) This prevents motion detection from triggering in those regions:(2)

motion:
  mask:
    - 0,0,100,0,100,100,0,100

(2)

The best approach is to start by increasing the threshold and min_score for the rabbit object, then fine-tune motion detection settings if needed.(1)(2)

📚 Sources:

1. Filters
2. Tuning Motion Detection

---

Have another question? Just tag @inkeep.

[hawkeye217]

The AI's response is what I would suggest, but this will be a little tougher until you can train more true positives and the base model improves. You'll want to aim for a ratio of 80/20 of true to false positives.