Face Recognition: VBR vs CBR (examples)

[kartalsmart]

CBR:

vs

VBR

Thought it might be useful, the difference is staggering. This is with 2688×1520 res cameras, min facesize 2000px. The system is working brilliantly with Frigate+ yolov9s model.

[inkeep-ai-assistant[bot]]

The provided documentation doesn't specifically discuss the impact of VBR (Variable Bit Rate) vs CBR (Constant Bit Rate) on face recognition performance. However, I can see from your examples that CBR provides significantly clearer face images compared to VBR, which would directly impact Frigate's face recognition accuracy.

For face recognition configuration in Frigate, you can adjust several parameters that may help optimize performance with different encoding types (1):

face_recognition:
  enabled: True
  min_area: 500
  detection_threshold: 0.7
  recognition_threshold: 0.9
  blur_confidence_filter: True

(1)

The blur_confidence_filter setting is particularly relevant to your observation, as it applies a blur quality filter to adjust confidence based on the blur level of the image (1). This could help compensate for the quality differences you're seeing between VBR and CBR encoding.

Your minimum face size of 2000px is much higher than the default min_area: 500 (1), which should help ensure only high-quality face detections are processed for recognition.

For more specific guidance on optimizing camera encoding settings for face recognition, you might want to check out these resources:

• License Plate Recognition (LPR)
• Supported Hardware
• Full Reference Config

Would you like a general answer about video encoding impacts on computer vision tasks?

📚 Sources:

1. Full Reference Config

---

Have another question? Just tag @inkeep.

[NickM-27]

Thanks! I'll look at including this in the docs now specifically

[cvroque]

Now I know why face recognition never worked properly here, I feel like throwing my Tapo cameras in the trash but at least I know what to look for in the future.