How to configure the python values to fix the facial movements of a new 3d character not originally thought to work with those

Question

I've just tracked and saved on the timeline the facial movements of Victor using OPEN CV and following this tutorial :

https://www.youtube.com/watch?v=RY_eErKlilw

The result has been good,as u can see :

https://drive.google.com/open?id=1ErXgMbFdy_bc7przM0LNe1BX2ud0GAxA

What I did next has been to apply the armature of Victor to a new character because I wanted to achieve the same goal as before with a new body,but the result has not been good :

https://drive.google.com/open?id=159YpyOV5Hxmu6b2drKVhIi4YWiQgU7XJ

The mouth moves in a totally wrong way and also the eyebrows. This happens because I'm not able to change the values and the maths formulas used on the python code that has been used to create the facial movements of Victor. Below you can read the code :

import bpy
import cv2
import time
import numpy

# Download trained model (lbfmodel.yaml)
# https://github.com/kurnianggoro/GSOC2017/tree/master/data

# Install prerequisites:

# Linux: (may vary between distro's and installation methods)
# This is for manjaro with Blender installed from the package manager
# python3 -m ensurepip
# python3 -m pip install --upgrade pip --user
# python3 -m pip install opencv-contrib-python numpy --user

# MacOS
# open the Terminal
# cd /Applications/Blender.app/Contents/Resources/2.81/python/bin
# ./python3.7m -m ensurepip
# ./python3.7m -m pip install --upgrade pip --user
# ./python3.7m -m pip install opencv-contrib-python numpy --user

# Windows:
# Open Command Prompt as Administrator
# cd "C:\Program Files\Blender Foundation\Blender 2.81\2.81\python\bin"
# python -m pip install --upgrade pip
# python -m pip install opencv-contrib-python numpy

class OpenCVAnimOperator(bpy.types.Operator):
    """Operator which runs its self from a timer"""
    bl_idname = "wm.opencv_operator"
    bl_label = "OpenCV Animation Operator"

    # Set paths to trained models downloaded above
    face_detect_path = cv2.data.haarcascades + "haarcascade_frontalface_default.xml"
    #landmark_model_path = "/home/username/Documents/Vincent/lbfmodel.yaml"  #Linux
    #landmark_model_path = "/Users/username/Downloads/lbfmodel.yaml"         #Mac
    landmark_model_path = "C:\\Users\\marietto2020\\Desktop\\OpenCV\\lbfmodel.yaml"    #Windows

    # Load models
    fm = cv2.face.createFacemarkLBF()
    fm.loadModel(landmark_model_path)
    cas = cv2.CascadeClassifier(face_detect_path)

    _timer = None
    _cap  = None
    stop = False

    # Webcam resolution:
    width = 640
    height = 480

    # 3D model points. 
    model_points = numpy.array([
                                (0.0, 0.0, 0.0),             # Nose tip
                                (0.0, -330.0, -65.0),        # Chin
                                (-225.0, 170.0, -135.0),     # Left eye left corner
                                (225.0, 170.0, -135.0),      # Right eye right corne
                                (-150.0, -150.0, -125.0),    # Left Mouth corner
                                (150.0, -150.0, -125.0)      # Right mouth corner
                            ], dtype = numpy.float32)
    # Camera internals
    camera_matrix = numpy.array(
                            [[height, 0.0, width/2],
                            [0.0, height, height/2],
                            [0.0, 0.0, 1.0]], dtype = numpy.float32
                            )

    # Keeps a moving average of given length
    def smooth_value(self, name, length, value):
        if not hasattr(self, 'smooth'):
            self.smooth = {}
        if not name in self.smooth:
            self.smooth[name] = numpy.array([value])
        else:
            self.smooth[name] = numpy.insert(arr=self.smooth[name], obj=0, values=value)
            if self.smooth[name].size > length:
                self.smooth[name] = numpy.delete(self.smooth[name], self.smooth[name].size-1, 0)
        sum = 0
        for val in self.smooth[name]:
            sum += val
        return sum / self.smooth[name].size

    # Keeps min and max values, then returns the value in a range 0 - 1
    def get_range(self, name, value):
        if not hasattr(self, 'range'):
            self.range = {}
        if not name in self.range:
            self.range[name] = numpy.array([value, value])
        else:
            self.range[name] = numpy.array([min(value, self.range[name][0]), max(value, self.range[name][1])] )
        val_range = self.range[name][1] - self.range[name][0]
        if val_range != 0:
            return (value - self.range[name][0]) / val_range
        else:
            return 0.0

    # The main "loop"
    def modal(self, context, event):

        if (event.type in {'RIGHTMOUSE', 'ESC'}) or self.stop == True:
            self.cancel(context)
            return {'CANCELLED'}

        if event.type == 'TIMER':
            self.init_camera()
            _, image = self._cap.read()
            #gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
            #gray = cv2.equalizeHist(gray)

            # find faces
            faces = self.cas.detectMultiScale(image, 
                scaleFactor=1.05,  
                minNeighbors=3, 
                flags=cv2.CASCADE_SCALE_IMAGE, 
                minSize=(int(self.width/5), int(self.width/5)))

            #find biggest face, and only keep it
            if type(faces) is numpy.ndarray and faces.size > 0: 
                biggestFace = numpy.zeros(shape=(1,4))
                for face in faces:
                    if face[2] > biggestFace[0][2]:
                        print(face)
                        biggestFace[0] = face

                # find the landmarks.
                _, landmarks = self.fm.fit(image, faces=biggestFace)
                for mark in landmarks:
                    shape = mark[0]

                    #2D image points. If you change the image, you need to change vector
                    image_points = numpy.array([shape[30],     # Nose tip - 31
                                                shape[8],      # Chin - 9
                                                shape[36],     # Left eye left corner - 37
                                                shape[45],     # Right eye right corne - 46
                                                shape[48],     # Left Mouth corner - 49
                                                shape[54]      # Right mouth corner - 55
                                            ], dtype = numpy.float32)

                    dist_coeffs = numpy.zeros((4,1)) # Assuming no lens distortion

                    # determine head rotation
                    if hasattr(self, 'rotation_vector'):
                        (success, self.rotation_vector, self.translation_vector) = cv2.solvePnP(self.model_points, 
                            image_points, self.camera_matrix, dist_coeffs, flags=cv2.SOLVEPNP_ITERATIVE, 
                            rvec=self.rotation_vector, tvec=self.translation_vector, 
                            useExtrinsicGuess=True)
                    else:
                        (success, self.rotation_vector, self.translation_vector) = cv2.solvePnP(self.model_points, 
                            image_points, self.camera_matrix, dist_coeffs, flags=cv2.SOLVEPNP_ITERATIVE, 
                            useExtrinsicGuess=False)

                    if not hasattr(self, 'first_angle'):
                        self.first_angle = numpy.copy(self.rotation_vector)

                    # set bone rotation/positions
                    bones = bpy.data.objects["RIG-Vincent"].pose.bones

                    # head rotation 
                    bones["head_fk"].rotation_euler[0] = self.smooth_value("h_x", 5, (self.rotation_vector[0] - self.first_angle[0])) / 1   # Up/Down
                    bones["head_fk"].rotation_euler[2] = self.smooth_value("h_y", 5, -(self.rotation_vector[1] - self.first_angle[1])) / 1.5  # Rotate
                    bones["head_fk"].rotation_euler[1] = self.smooth_value("h_z", 5, (self.rotation_vector[2] - self.first_angle[2])) / 1.3   # Left/Right

                    bones["head_fk"].keyframe_insert(data_path="rotation_euler", index=-1)

                    # mouth position
                    bones["mouth_ctrl"].location[2] = self.smooth_value("m_h", 2, -self.get_range("mouth_height", numpy.linalg.norm(shape[62] - shape[66])) * 0.06 )
                    bones["mouth_ctrl"].location[0] = self.smooth_value("m_w", 2, (self.get_range("mouth_width", numpy.linalg.norm(shape[54] - shape[48])) - 0.5) * -0.04)

                    bones["mouth_ctrl"].keyframe_insert(data_path="location", index=-1)

                    #eyebrows
                    bones["brow_ctrl_L"].location[2] = self.smooth_value("b_l", 3, (self.get_range("brow_left", numpy.linalg.norm(shape[19] - shape[27])) -0.5) * 0.04)
                    bones["brow_ctrl_R"].location[2] = self.smooth_value("b_r", 3, (self.get_range("brow_right", numpy.linalg.norm(shape[24] - shape[27])) -0.5) * 0.04)

                    bones["brow_ctrl_L"].keyframe_insert(data_path="location", index=2)
                    bones["brow_ctrl_R"].keyframe_insert(data_path="location", index=2)

                    # eyelids
                    l_open = self.smooth_value("e_l", 2, self.get_range("l_open", -numpy.linalg.norm(shape[48] - shape[44]))  )
                    r_open = self.smooth_value("e_r", 2, self.get_range("r_open", -numpy.linalg.norm(shape[41] - shape[39]))  )
                    eyes_open = (l_open + r_open) / 2.0 # looks weird if both eyes aren't the same...
                    bones["eyelid_up_ctrl_R"].location[2] =   -eyes_open * 0.025 + 0.005
                    bones["eyelid_low_ctrl_R"].location[2] =  eyes_open * 0.025 - 0.005
                    bones["eyelid_up_ctrl_L"].location[2] =   -eyes_open * 0.025 + 0.005
                    bones["eyelid_low_ctrl_L"].location[2] =  eyes_open * 0.025 - 0.005

                    bones["eyelid_up_ctrl_R"].keyframe_insert(data_path="location", index=2)
                    bones["eyelid_low_ctrl_R"].keyframe_insert(data_path="location", index=2)
                    bones["eyelid_up_ctrl_L"].keyframe_insert(data_path="location", index=2)
                    bones["eyelid_low_ctrl_L"].keyframe_insert(data_path="location", index=2)

                    # draw face markers
                    for (x, y) in shape:
                        cv2.circle(image, (x, y), 2, (0, 255, 255), -1)

            # draw detected face
            for (x,y,w,h) in faces:
                cv2.rectangle(image,(x,y),(x+w,y+h),(255,0,0),1)

            # Show camera image in a window                     
            cv2.imshow("Output", image)
            cv2.waitKey(1)

        return {'PASS_THROUGH'}

    def init_camera(self):
        if self._cap == None:
            self._cap = cv2.VideoCapture(0)
            self._cap.set(cv2.CAP_PROP_FRAME_WIDTH, self.width)
            self._cap.set(cv2.CAP_PROP_FRAME_HEIGHT, self.height)
            self._cap.set(cv2.CAP_PROP_BUFFERSIZE, 1)
            time.sleep(1.0)

    def stop_playback(self, scene):
        print(format(scene.frame_current) + " / " + format(scene.frame_end))
        if scene.frame_current == scene.frame_end:
            bpy.ops.screen.animation_cancel(restore_frame=False)

    def execute(self, context):
        bpy.app.handlers.frame_change_pre.append(self.stop_playback)

        wm = context.window_manager
        self._timer = wm.event_timer_add(0.01, window=context.window)
        wm.modal_handler_add(self)
        return {'RUNNING_MODAL'}

    def cancel(self, context):
        wm = context.window_manager
        wm.event_timer_remove(self._timer)
        cv2.destroyAllWindows()
        self._cap.release()
        self._cap = None

def register():
    bpy.utils.register_class(OpenCVAnimOperator)

def unregister():
    bpy.utils.unregister_class(OpenCVAnimOperator)

if __name__ == "__main__":
    register()

    # test call
    #bpy.ops.wm.opencv_operator()

I tried to change some value,but since I don't understand what piece of code makes what,I haven't been able to fix the movements. Can some one help me to do that ? thanks.

Answer 1

To understand what the numbers represents you need to consider the context in which they are stated.

If you wish to avoid coding - you may want to try and error. Modify values and observe the result. copy the script. Use the copy and don't register it. Change one value at the time.

Some of these values are not related to Blender but the OpenCV - the keymaps - in some situation you may want to adjust certain values, lists length, as they modify the target and may better suits your vision.

Values like : -eyes_open * 0.025 + 0.005 - set a limit to the opening of the eye

The 0,2, and occasionally 1 in the following: bones["mouth_ctrl"].location[2] = ...

are for the axis. 0,1,2 for x, y, z accordingly. Y is considered only for rotation, as the location source is a 2D with X and Y (Z in blender).

More effective approach may be to learn some coding syntax so to better understand your actions.

Whatever you decide, please note that this script is limited to start with. A great starting point but a work in progress.

To really adapt the model to your animation vision, you need to add more code lines for the movement of the mouth and the eyes.

Ex: check # mouth position the mouth bones only consider the upper and lower lips OpenCV keypoints, whereas 'real' mouth animation tweaks the corners, narrow and widen, a-symmetrical movements, cheeks deformation (for which you have not source keypoints on OpenCV) and also rotation on the X axis, which you may get with additional cameras and some additional coding.

Similar considerations apply to the eyes and nose movements

I know you wanted a operational solution. Sorry I don't have a magic stick. I hope it can help you keep the exploration.