Haar Cascades: Limiting detectMultiScale to just one set of facial features?

asked 2018-12-29 19:25:06 -0600

byoon
21 ●2

I am currently running OpenCV version 3.4.4 on Python 3.7.1 and designed a program that utilizes the eye haar cascade classifiers. The program runs fine but often misinterprets the center lower lip or one of the nostrils as a false eye. Is this a problem with the cascade I am using? Are there ways to limit my program to only the two most prominent eye features?

Here is an example: image description

While on the subject of Haar Cascades, when using the detectMultiScale function to find the position of the face, I noticed that when the face is moved upwards, the y-value actually decreases. Why might that be the case?

Any and all help is appreciated, B

edit retag flag offensive close merge delete

add a comment

1 answer

Sort by » oldest newest most voted

answered 2018-12-29 22:29:50 -0600

berak
32993 ●7 ●81 ●312

updated 2018-12-29 23:16:24 -0600

misinterprets the center lower lip or one of the nostrils as a false eye.

indeed, the eye / mouth cascades are quite inaccurate. there are better alternatives, like facial landmarks (e.g. from dlib, flandmark, openpose. opencv also has pretrained landmarks models, however, they're not usable from python, currently)