Adding new plugins (Extract_Align & Convert_Masked)

.gitignore

@@ -4,6 +4,6 @@
 !*.txt
 !Dockerfile*
 !requirements*
-!lib/*.py
+!lib
 !scripts
 !plugins

lib/aligner.py

@@ -0,0 +1,26 @@
+import numpy
+
+from lib.umeyama import umeyama
+
+mean_face_x = numpy.array([
+0.000213256, 0.0752622, 0.18113, 0.29077, 0.393397, 0.586856, 0.689483, 0.799124,
+0.904991, 0.98004, 0.490127, 0.490127, 0.490127, 0.490127, 0.36688, 0.426036,
+0.490127, 0.554217, 0.613373, 0.121737, 0.187122, 0.265825, 0.334606, 0.260918,
+0.182743, 0.645647, 0.714428, 0.793132, 0.858516, 0.79751, 0.719335, 0.254149,
+0.340985, 0.428858, 0.490127, 0.551395, 0.639268, 0.726104, 0.642159, 0.556721,
+0.490127, 0.423532, 0.338094, 0.290379, 0.428096, 0.490127, 0.552157, 0.689874,
+0.553364, 0.490127, 0.42689 ])
+
+mean_face_y = numpy.array([
+0.106454, 0.038915, 0.0187482, 0.0344891, 0.0773906, 0.0773906, 0.0344891,
+0.0187482, 0.038915, 0.106454, 0.203352, 0.307009, 0.409805, 0.515625, 0.587326,
+0.609345, 0.628106, 0.609345, 0.587326, 0.216423, 0.178758, 0.179852, 0.231733,
+0.245099, 0.244077, 0.231733, 0.179852, 0.178758, 0.216423, 0.244077, 0.245099,
+0.780233, 0.745405, 0.727388, 0.742578, 0.727388, 0.745405, 0.780233, 0.864805,
+0.902192, 0.909281, 0.902192, 0.864805, 0.784792, 0.778746, 0.785343, 0.778746,
+0.784792, 0.824182, 0.831803, 0.824182 ])
+
+landmarks_2D = numpy.stack( [ mean_face_x, mean_face_y ], axis=1 )
+
+def get_align_mat(face):
+    return umeyama( numpy.array(face.landmarks[17:]), landmarks_2D, True )[0:2]

lib/faces_detect.py

@@ -1,13 +1,28 @@
 import face_recognition
+import face_recognition_models
+import dlib
 from .DetectedFace import DetectedFace
 
 def detect_faces(frame):
     face_locations = face_recognition.face_locations(frame)
-    #face_encodings = face_recognition.face_encodings(frame, face_locations)
+    landmarks = _raw_face_landmarks(frame, face_locations)
+    landmarks_as_tuples = [[(p.x, p.y) for p in landmark.parts()] for landmark in landmarks]
 
-    for (top, right, bottom, left) in face_locations:
+    for ((top, right, bottom, left), landmarks) in zip(face_locations, landmarks_as_tuples):
         x = left
         y = top
         w = right - left
         h = bottom - top
-        yield DetectedFace(frame[y: y + h, x: x + w], x, w, y, h, None)
+        yield DetectedFace(frame[y: y + h, x: x + w], x, w, y, h, landmarks)
+
+# Copy/Paste (mostly) from private method in face_recognition
+predictor_68_point_model = face_recognition_models.pose_predictor_model_location()
+pose_predictor = dlib.shape_predictor(predictor_68_point_model)
+
+def _raw_face_landmarks(face_image, face_locations):
+    face_locations = [_css_to_rect(face_location) for face_location in face_locations]
+    return [pose_predictor(face_image, face_location) for face_location in face_locations]
+
+def _css_to_rect(css):
+    return dlib.rectangle(css[3], css[0], css[1], css[2])
+# end of Copy/Paste

lib/utils.py

@@ -12,7 +12,7 @@ def get_folder(path):
 
 
 def get_image_paths(directory):
-    return [x.path for x in scandir(directory) if x.name.endswith('.jpg') or x.name.endswith('.png')]
+    return [x.path for x in scandir(directory) if x.name.endswith('.jpg') or x.name.endswith('.jpeg') or x.name.endswith('.png')]
 
 
 def load_images(image_paths, convert=None):

plugins/Convert_Adjust.py

@@ -2,28 +2,16 @@ import cv2
 import numpy
 import os
 
-from lib.model import autoencoder_A
-from lib.model import autoencoder_B
-from lib.model import encoder, decoder_A, decoder_B
-
 class Convert(object):
-    def __init__(self,
-                model_dir="models",
-                swap_model=False,
-                use_aligner=False):
-        face_A = '/decoder_A.h5' if not swap_model else '/decoder_B.h5'
-        face_B = '/decoder_B.h5' if not swap_model else '/decoder_A.h5'
+    def __init__(self, autoencoder):
+        self.autoencoder = autoencoder
+         
+        self.use_smooth_mask=True
+        self.use_avg_color_adjust=True
 
-        encoder.load_weights(model_dir + "/encoder.h5")
-        decoder_A.load_weights(model_dir + face_A)
-        decoder_B.load_weights(model_dir + face_B)
-
-        self.autoencoder = autoencoder_B
-
-    def convert_one_image(self,image, 
-                        use_smooth_mask=True,
-                        use_avg_color_adjust=True):
-        assert image.shape == (256, 256, 3)
+    def convert_one_image( self, original, face_detected ):
+        #assert image.shape == (256, 256, 3)
+        image = cv2.resize(face_detected.image, (256, 256))
         crop = slice(48, 208)
         face = image[crop, crop]
         old_face = face.copy()
@@ -34,12 +22,14 @@ class Convert(object):
         new_face = numpy.clip(new_face * 255, 0, 255).astype(image.dtype)
         new_face = cv2.resize(new_face, (160, 160))
 
-        if use_avg_color_adjust:
+        if self.use_avg_color_adjust:
             self.adjust_avg_color(old_face,new_face)
-        if use_smooth_mask:
+        if self.use_smooth_mask:
             self.smooth_mask(old_face,new_face)
 
-        return self.superpose(image, new_face, crop)
+        new_face = self.superpose(image, new_face, crop)
+        original[slice(face_detected.y, face_detected.y + face_detected.h), slice(face_detected.x, face_detected.x + face_detected.w)] = cv2.resize(new_face, (face_detected.w, face_detected.h))
+        return original
 
     def adjust_avg_color(self,img_old,img_new):
         w,h,c = img_new.shape

plugins/Convert_Masked.py

@@ -0,0 +1,81 @@
+import cv2
+import numpy
+
+from lib.aligner import get_align_mat
+
+# Based on: https://gist.github.com/anonymous/d3815aba83a8f79779451262599b0955
+
+class Convert():
+    def __init__(self, autoencoder):
+        self.autoencoder = autoencoder
+
+        # if args.erosionKernelSize > 0:
+        #     self.erosion_kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE,(args.erosionKernelSize,args.erosionKernelSize))
+        # else:
+        #     self.erosion_kernel = None
+
+        self.erosion_kernel = None
+        self.blurSize = 2
+        self.seamlessClone = False
+        self.maskType = 'Rect' # Choose in 'FaceHullAndRect','FaceHull','Rect'
+
+    def convert_one_image( self, image, face_detected ):
+        size = 64
+        image_size = image.shape[1], image.shape[0]
+
+        mat = numpy.array(get_align_mat(face_detected)).reshape(2,3) * size
+
+        new_face = self.get_new_face(image,mat,size)
+
+        image_mask = self.get_image_mask( image, new_face, face_detected, mat, image_size )
+
+        base_image = numpy.copy( image )
+        new_image = numpy.copy( image )
+
+        cv2.warpAffine( new_face, mat, image_size, new_image, cv2.WARP_INVERSE_MAP, cv2.BORDER_TRANSPARENT )
+
+        outImage = None
+        if self.seamlessClone:
+            masky,maskx = cv2.transform( numpy.array([ size/2,size/2 ]).reshape(1,1,2) ,cv2.invertAffineTransform(mat) ).reshape(2).astype(int)
+            outimage = cv2.seamlessClone(new_image.astype(numpy.uint8),base_image.astype(numpy.uint8),(image_mask*255).astype(numpy.uint8),(masky,maskx) , cv2.NORMAL_CLONE )
+        else:
+            foreground = cv2.multiply(image_mask, new_image.astype(float))
+            background = cv2.multiply(1.0 - image_mask, base_image.astype(float))
+            outimage = cv2.add(foreground, background)
+
+        return outimage
+
+    def get_new_face(self, image, mat, size):
+        face = cv2.warpAffine( image, mat, (size,size) )
+        face = numpy.expand_dims( face, 0 )
+        new_face = self.autoencoder.predict( face / 255.0 )[0]
+
+        return numpy.clip( new_face * 255, 0, 255 ).astype( image.dtype )
+
+    def get_image_mask(self, image, new_face, face_detected, mat, image_size):
+
+        face_mask = numpy.zeros(image.shape,dtype=float)
+        if 'Rect' in self.maskType:
+            face_src = numpy.ones(new_face.shape,dtype=float) 
+            cv2.warpAffine( face_src, mat, image_size, face_mask, cv2.WARP_INVERSE_MAP, cv2.BORDER_TRANSPARENT )
+
+        hull_mask = numpy.zeros(image.shape,dtype=float)
+        if 'Hull' in self.maskType:
+            hull = cv2.convexHull( numpy.array( face_detected.landmarks ).reshape((-1,2)).astype(int) ).flatten().reshape( (-1,2) )
+            cv2.fillConvexPoly( hull_mask,hull,(1,1,1) )
+
+        if self.maskType == 'Rect':
+            image_mask = face_mask
+        elif self.maskType == 'FaceHull':
+            image_mask = hull_mask
+        else:
+            image_mask = ((face_mask*hull_mask))
+
+
+        if self.erosion_kernel is not None:
+            image_mask = cv2.erode(image_mask,self.erosion_kernel,iterations = 1)
+
+        if self.blurSize!=0:
+            image_mask = cv2.blur(image_mask,(self.blurSize,self.blurSize))
+
+        return image_mask

plugins/Extract_Align.py

@@ -0,0 +1,16 @@
+import cv2
+
+from lib.aligner import get_align_mat
+
+class Extract(object):
+    def extract(self, image, face, size):
+        if face.landmarks == None:
+            return cv2.resize(face.image, (size, size))
+
+        alignment = get_align_mat( face )
+        return self.transform( image, alignment, size, 48 )
+    
+    def transform( self, image, mat, size, padding=0 ):
+        mat = mat * (size - 2 * padding)
+        mat[:,2] += padding
+        return cv2.warpAffine( image, mat, ( size, size ) )

plugins/Extract_Crop.py

@@ -2,4 +2,4 @@ import cv2
 
 class Extract(object):
     def extract(self, image, face, size):
-        return cv2.resize(face.image, size)
+        return cv2.resize(face.image, (size, size))

scripts/convert.py

@@ -1,9 +1,15 @@
 import cv2
-from lib.cli import DirectoryProcessor, FullPaths
+
 from pathlib import Path
+from lib.cli import DirectoryProcessor, FullPaths
 from lib.faces_detect import detect_faces
 
-from plugins.Convert_Adjust import Convert
+from lib.model import autoencoder_A
+from lib.model import autoencoder_B
+from lib.model import encoder, decoder_A, decoder_B
+
+#from plugins.Convert_Adjust import Convert
+from plugins.Convert_Masked import Convert
 
 class ConvertImage(DirectoryProcessor):
     filename = ''
@@ -31,20 +37,29 @@ class ConvertImage(DirectoryProcessor):
         return parser
 
     def process_image(self, filename):
+        # TODO move the model load and the converter creation in a method called on init, but after the arg parsing
+        (face_A,face_B) = ('/decoder_A.h5', '/decoder_B.h5') if not self.arguments.swap_model else ('/decoder_B.h5', '/decoder_A.h5')
+
+        model_dir = self.arguments.model_dir
+        encoder.load_weights(model_dir + "/encoder.h5")
+        decoder_A.load_weights(model_dir + face_A)
+        decoder_B.load_weights(model_dir + face_B)
+
+        converter = Convert(autoencoder_B)
+
         try:
-            converter = Convert(self.arguments.model_dir, self.arguments.swap_model)
             image = cv2.imread(filename)
             for (idx, face) in enumerate(detect_faces(image)):
                 if idx > 0 and self.arguments.verbose:
                     print('- Found more than one face!')
                     self.verify_output = True
 
-                new_face = converter.convert_one_image(cv2.resize(face.image, (256, 256)))
-                image[slice(face.y, face.y + face.h), slice(face.x, face.x + face.w)] = cv2.resize(new_face, (face.w, face.h))
+                image = converter.convert_one_image(image, face)
                 self.faces_detected = self.faces_detected + 1
+
             output_file = self.output_dir / Path(filename).name
             cv2.imwrite(str(output_file), image)
         except Exception as e:
-            print('Failed to extract from image: {}. Reason: {}'.format(filename, e))
+            print('Failed to convert image: {}. Reason: {}'.format(filename, e))
 
 

scripts/extract.py

@@ -3,7 +3,7 @@ import cv2
 from pathlib import Path
 from lib.cli import DirectoryProcessor
 from lib.faces_detect import detect_faces
-from plugins.Extract_Crop import Extract
+from plugins.Extract_Align import Extract
 
 class ExtractTrainingData(DirectoryProcessor):
     def create_parser(self, subparser, command, description):
@@ -24,7 +24,7 @@ class ExtractTrainingData(DirectoryProcessor):
                     print('- Found more than one face!')
                     self.verify_output = True
 
-                resized_image = extractor.extract(image, face, (256, 256))
+                resized_image = extractor.extract(image, face, 256)
                 output_file = self.output_dir / Path(filename).stem
                 cv2.imwrite(str(output_file) + str(idx) + Path(filename).suffix, resized_image)
                 self.faces_detected = self.faces_detected + 1