Python – Viola-Jones in Python with openCV, detection mouth and nose


I have an algorithm Viola-Jones in Python. I'm using haarcascade xml, which I load from openCV root file. But there wasn't any xml file for mouth and nose in openCV, so I downloaded these files from EmguCV. Result for detection of face is OK, but detection of eye isn't good and nose with mouth is very bad. I tried to change parameters in face_cascade.detectMultiScale, but it didn't help at all.

My code:

import cv2
import sys

def facedet(img):
    face_cascade = cv2.CascadeClassifier('/home/kattynka/opencv/data/haarcascades/haarcascade_frontalface_default.xml')
    eye_cascade = cv2.CascadeClassifier('/home/kattynka/opencv/data/haarcascades/haarcascade_eye.xml')
    mouth_cascade = cv2.CascadeClassifier('/home/kattynka/opencv/data/haarcascades/haarcascade_mcs_mouth.xml')
    nose_cascade = cv2.CascadeClassifier('/home/kattynka/opencv/data/haarcascades/haarcascade_mcs_nose.xml')

    img = cv2.imread(img)
    gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)

    faces = face_cascade.detectMultiScale(gray, 1.3, 5)

    for (x,y,w,h) in faces:
        cv2.rectangle(img, (x,y), (x+w,y+h), (255,0,0), 2)
        roi_gray = gray[y:y+h, x:x+w]
        roi_color = img[y:y+h, x:x+w]
        eyes = eye_cascade.detectMultiScale(roi_gray)
        nose =  nose_cascade.detectMultiScale(roi_gray)
        mouth = mouth_cascade.detectMultiScale(roi_gray)

        for (ex,ey,ew,eh) in eyes:
            cv2.rectangle(roi_color, (ex,ey), (ex+ew, ey+eh), (0,255,0), 2)
        for (nx, ny, nw, nh) in nose:
            cv2.rectangle(roi_color, (nx, ny), (nx + nw, ny + nh), (0, 0, 255), 2)
        for (mx, my, mw, mh) in mouth:
            cv2.rectangle(roi_color, (mx, my), (mx + mw, my + mh), (0, 0, 0), 2)

    cv2.namedWindow('image', cv2.WINDOW_NORMAL)

if __name__ == '__main__':
    #img = sys.argv[1]

My question

What am I doing wrong? Is there any simple solution, which will give me a better result?


enter image description here

enter image description here

enter image description here

enter image description here

Best Solution

Haar cascades perform alright for faces but not so well for smaller individual parts. A better solution is to detect all the face landmarks together. A good algorithm for that is "One Millisecond Face Alignment with an Ensemble of Regression Trees by Vahid Kazemi and Josephine Sullivan, CVPR 2014" which is implemented in Dlib (