Low confidence score (~0.3) for clear face images using SCRFD model

[satya-gopal]

I have integrated the SCRFD face detection model into my application. However, I am encountering an issue where the model returns a very low confidence score (~0.3) even when the input image contains a clearly visible and well-defined face.

Details:
The input image has:
A clear, frontal face
Good lighting conditions
No significant occlusion
Expected [behavior:]
The model should return a higher confidence score for such images
Actual behavior:
The model outputs a confidence score around 0.3, which is unexpectedly low

Could you please help clarify:

What could cause such low confidence scores for clear face images?
Are there recommended preprocessing steps or thresholds for optimal results?

[rohitrrr]

SCRFD fails on full-frame faces because the face exceeds the receptive field of the model's largest anchors, meaning the feature maps can no longer "see" the face as a single, cohesive object.

[abhay288]

A possible workaround could be to resize the input image so that the face occupies a smaller portion of the frame before running SCRFD. This may help improve the confidence score for very large faces.