Ask Your Question

Revision history [back]

click to hide/show revision 1
initial version

Interpreting OpenCV FaceRecognition predicted confidence values

I am toying with the various FaceRecognition algorithms, and I'd like to better understand the confidence values so that I can have a sense of when to ignore a match or when I can rely on a match.

Using the ATT face database, I did test 1 where I trained on the 40 faces then ran prediction on a known face (With unknown image of course). I then did a second test where I trained on 39 faces and ran prediction on an unknown face (happens to be the same image as used in test 1).

The values I got were: Eigenspace - 1806 when face known - 2618 when face unknown

Fisherface - 372 known - 841 unknown

LBPH - 36 known - 55 unknown.

If I am interpreting the algorithms correctly, Eigenspace and Fisherface work in a high dimension space and try to find the closest neighbor for a given test image. This means the confidence value will change depending on data set, and I can not have a simple threshold. Is there any other information I can gather, such as average distance between clusters, so that I can understand if I should keep or ignore a prediction?

In regards to LBPH, is this confidence acting the same way?

Many thanks

Interpreting OpenCV FaceRecognition predicted confidence values

I am toying with the various FaceRecognition algorithms, and I'd like to better understand the confidence values so that I can have a sense of when to ignore a match or when I can rely on a match.



Using the ATT face database, I did test 1 where I trained on the 40 faces then ran prediction on a known face (With unknown image of course). I then did a second test where I trained on 39 faces and ran prediction on an unknown face (happens to be the same image as used in test 1).



The values I got were: Eigenspace were:
Eigenspace
- 1806 when face known known
- 2618 when face unknown

Fisherface

Fisherface
- 372 known known
- 841 unknown

LBPH

LBPH
- 36 known known
- 55 unknown.

unknown

If I am interpreting the algorithms correctly, Eigenspace and Fisherface work in a high dimension space and try to find the closest neighbor for a given test image. This means the confidence value will change depending on data set, and I can not have a simple threshold. Is there any other information I can gather, such as average distance between clusters, so that I can understand if I should keep or ignore a prediction?



In regards to LBPH, is this confidence acting the same way?



Many thanks