Ask Your Question
0

peak in false positives for detectMultiscale

asked 2020-02-17 03:48:03 -0500

LangdalePikes gravatar image

updated 2020-02-17 06:30:49 -0500

I am doing a little research into the optimum detectMultiScale parameters to use for a face identification project and have plotted the number of false Positives against scale factor for min Neighbors 3,4,5,6, across minimum feature sizes 20:20, 30:30, and 40:40. See the attached graph.

I am using lbpcascade_frontalface.xml as the cascade classifier which was trained with a height and width of 24x24.

image description

And I am using greyscale images resized to 200x200.

My question is why is there a an obvious and large peak in the number of false Positives at aroundscale Factor 1.45 (1.35 to 1.65) for each plot?

image description

Thanks in anticipation of your help.

Langdale

edit retag flag offensive close merge delete

Comments

that range of scale factors spans values that I would not recommend because they're too large and may miss detections (false negative). as for why you get false positives... no idea. try a denser and uniform sampling of scale factors. I would guess that the values near 1.05 and 1.45 are proper responses and the valleys beside those are how the algorithm fails catastrophically because it entirely misses the scales at which there are faces in your test images.

crackwitz gravatar imagecrackwitz ( 2020-02-19 09:39:53 -0500 )edit

1 answer

Sort by ยป oldest newest most voted
0

answered 2020-03-29 05:11:02 -0500

ComputerVisionary gravatar image

The LBP, HAAR, HOG detector still detect faces even at "wrong" scales due to the great capabilities of the model to fit something similar to a "face" in "tricky" conditions. You are probably detecting faces at a wrong scale and the fine tuning of the position of the face (or sub-parts of the faces) fails.

Have your tried to print the detected faces at those scales?

For an expert point of view about the opencv cascades with opencv 2, 3, and 4 take a look here: http://www.vision-ary.net/2015/03/lar...

Regards!

edit flag offensive delete link more
Login/Signup to Answer

Question Tools

1 follower

Stats

Asked: 2020-02-17 03:48:03 -0500

Seen: 29 times

Last updated: 20 hours ago