How to set face detection parameters in OpenCV?

Question

OpenCV

There was such a problem that a face recognition script was written for the webcam, and the same script does not work for the ip camera ( it works, but does not detect faces ).
The first steps to solve this problem is to change the resolution, as it was 640x480p on the webcam, and 1280x720p or fullHD on the ip camera.
there is not much difference, all the same, the window opens in a large window, which even does not fit on the monitor.
read about faceCascade.detectMultiScale, and that there are several parameters such as scaleFactor, minNeighbors, minSize. From this I could understand that for the cascade you can increase something, you can adjust the distance between adjacent elements, the minimum window for face recognition.

The source code itself:

cam = cv2.VideoCapture("rtsp://имя юзера@ip.add.re.ss/mpeg4?username=имя юзера&password=пароль в MD5") #redacted

# cam.set(3, 640) # set video widht
# cam.set(4, 480) # set video height

cam.set(3, 1280) # set video widht
cam.set(4, 720) # set video height

state = True

is_start = open(projectpath+"/state.log")
state_for_start = is_start.readline()
if state_for_start == "0":
    state = True
    print("Запускаем скрипт")
else:
    state = False
    print("отредактируйте state.log")
is_start.close()

while state:

    is_start = open(projectpath+"/state.log")
    state_for_start = is_start.readline()
    if state_for_start == "0":
        state = True
#         print("Запускаем скрипт")
    else:
        state = False
        print("отредактируйте state.log")
    is_start.close()


    ret, img =cam.read()
    gray = cv2.cvtColor(img,cv2.COLOR_BGR2GRAY)
    faces = faceCascade.detectMultiScale( 
        gray,
        scaleFactor = 1.2,
        minNeighbors = 5, # origin
        minSize = (100, 100) # not origin
       )

    for(x,y,w,h) in faces:
        cv2.rectangle(img, (x,y), (x+w,y+h), (0,255,0), 2)
        id, confidence = recognizer.predict(gray[y:y+h,x:x+w])
 
        if (confidence < 100):
            id = names[id]
            confidence = "  {0}%".format(round(100 - confidence))
        else:
            id = "unknown"
            confidence = "  {0}%".format(round(100 - confidence))
            # ---> можно тыкнуть функцию сюда
        cv2.putText(img, str(id), (x+5,y-5), font, 1, (255,255,255), 2)
        cv2.putText(img, str(confidence), (x+5,y+h-5), font, 1, (255,255,0), 1)

#              Самообучение
# ------------------------------------------------------------------------------------------------
        if True:
            if id not in enterings:
                enterings.append(id)

            if id not in names:
                face_id=len(names)
                count += 1
                cv2.imwrite(path+"/User." + str(face_id) + '.' + str(count) + ".jpg", gray[y:y+h,x:x+w])
                if count > count_of_shots:
                    restudying()
                    names.append("User."+str(face_id))
                    recognizer.read(projectpath+'/trainer/trainer.yml')
                    f = open(projectpath+"/enterings.log", "a")
                    count=0
                    f.write(str(face_id)+" User"+str(face_id)+"\n")                    
                    f.close

                    with open(projectpath+"/names.log", "w") as file:
                        for name in names:
                            file.write(name + '\n')

    cv2.imshow('camera',img) 
    k = cv2.waitKey(10) & 0xff # Press 'ESC' for exiting video
    if k == 27:
        break
cam.release()
cv2.destroyAllWindows()

Some problems:
faces are not recognized at all, even if there is a person in the image,

something that should not be detected is detected

. Some minSize manipulations led to the fact that objects less than 50x50 are not recognized as a face



. What values ​​​​to select and what does it affect?

Answer 1

UPD: I realized that you can change the scaleFactor and get a better result, it would be great if someone gave a hint in which direction to google to put the background layer, and the neuron would not try to pass off parts of the landscape as a face