Update face alignment

app.py

@@ -4,17 +4,22 @@ os.system("pip -qq install facenet_pytorch")
 from facenet_pytorch import MTCNN
 from torchvision import transforms
 import torch, PIL
-from tqdm.notebook import tqdm
 import gradio as gr
 import torch
+from utils import align_face
 
 device = "cuda:0" if torch.cuda.is_available() else "cpu"
 
 image_size = 512
+transform_size = 1024
 
 means = [0.5, 0.5, 0.5]
 stds = [0.5, 0.5, 0.5]
 
+img_transforms = transforms.Compose([
+            transforms.ToTensor(),
+            transforms.Normalize(means, stds)])
+ 
 model_path = hf_hub_download(repo_id="jjeamin/ArcaneStyleTransfer", filename="pytorch_model.bin")
 
 if 'cuda' in device:
@@ -25,76 +30,14 @@ else:
     style_transfer = torch.jit.load(model_path).eval().cpu()
     t_stds = torch.tensor(stds).cpu()[:,None,None]
     t_means = torch.tensor(means).cpu()[:,None,None]
-    
-mtcnn = MTCNN(image_size=image_size, margin=80)
-
-def detect(img):
- 
-        # Detect faces
-        batch_boxes, batch_probs, batch_points = mtcnn.detect(img, landmarks=True)
-        # Select faces
-        if not mtcnn.keep_all:
-            batch_boxes, batch_probs, batch_points = mtcnn.select_boxes(
-                batch_boxes, batch_probs, batch_points, img, method=mtcnn.selection_method
-            )
- 
-        return batch_boxes, batch_points
-
-def makeEven(_x):
-  return _x if (_x % 2 == 0) else _x+1
-
-def scale(boxes, _img, max_res=1_500_000, target_face=256, fixed_ratio=0, max_upscale=2, VERBOSE=False):
- 
-    x, y = _img.size
- 
-    ratio = 2 #initial ratio
- 
-    #scale to desired face size
-    if (boxes is not None):
-      if len(boxes)>0:
-        ratio = target_face/max(boxes[0][2:]-boxes[0][:2]); 
-        ratio = min(ratio, max_upscale)
-        if VERBOSE: print('up by', ratio)
 
-    if fixed_ratio>0:
-      if VERBOSE: print('fixed ratio')
-      ratio = fixed_ratio
- 
-    x*=ratio
-    y*=ratio
- 
-    #downscale to fit into max res 
-    res = x*y
-    if res > max_res:
-      ratio = pow(res/max_res,1/2); 
-      if VERBOSE: print(ratio)
-      x=int(x/ratio)
-      y=int(y/ratio)
- 
-    #make dimensions even, because usually NNs fail on uneven dimensions due skip connection size mismatch
-    x = makeEven(int(x))
-    y = makeEven(int(y))
-    
-    size = (x, y)
-
-    return _img.resize(size)
-
-def scale_by_face_size(_img, max_res=1_500_000, target_face=256, fix_ratio=0, max_upscale=2, VERBOSE=False):
-    boxes = None
-    boxes, _ = detect(_img)
-    if VERBOSE: print('boxes',boxes)
-    img_resized = scale(boxes, _img, max_res, target_face, fix_ratio, max_upscale, VERBOSE)
-    return img_resized
-
-
-img_transforms = transforms.Compose([
-            transforms.ToTensor(),
-            transforms.Normalize(means, stds)])
- 
 def tensor2im(var):
      return var.mul(t_stds).add(t_means).mul(255.).clamp(0,255).permute(1,2,0)
 
 def proc_pil_img(input_image):
+    # input_image = PIL
+    input_image = align_face(input_image, output_size=image_size, transform_size=transform_size)
+    
     if 'cuda' in device: 
         transformed_image = img_transforms(input_image)[None,...].cuda().half()
     else:
@@ -108,7 +51,6 @@ def proc_pil_img(input_image):
     return output_image
 
 def process(im):
-    im = scale_by_face_size(im, target_face=image_size, max_res=1_500_000, max_upscale=1)
     res = proc_pil_img(im)
     return res
         

requirements.txt

@@ -7,3 +7,4 @@ scipy
 cmake
 onnxruntime-gpu
 opencv-python-headless
+dlib

utils.py

@@ -0,0 +1,114 @@
+import dlib
+import numpy as np
+import scipy
+from PIL import Image
+from huggingface_hub import hf_hub_download
+
+shape_predictor_path = hf_hub_download(repo_id="jjeamin/ArcaneStyleTransfer", filename="shape_predictor_68_face_landmarks.dat")
+
+def get_landmark(img, predictor):
+    """get landmark with dlib
+    :return: np.array shape=(68, 2)
+    """
+    detector = dlib.get_frontal_face_detector()
+
+    dets = detector(img, 1)
+    assert len(dets) > 0, "Face not detected, try another face image"
+
+    for k, d in enumerate(dets):
+        shape = predictor(img, d)
+
+    t = list(shape.parts())
+    a = []
+    for tt in t:
+        a.append([tt.x, tt.y])
+    lm = np.array(a)
+    return lm
+
+
+def align_face(img, output_size=512, transform_size=1024, enable_padding=True):
+
+    """
+    :param filepath: str
+    :return: PIL Image
+    """
+    np_img = np.array(img)
+    predictor = dlib.shape_predictor(shape_predictor_path)
+    lm = get_landmark(np_img, predictor)
+
+    lm_chin = lm[0: 17]  # left-right
+    lm_eyebrow_left = lm[17: 22]  # left-right
+    lm_eyebrow_right = lm[22: 27]  # left-right
+    lm_nose = lm[27: 31]  # top-down
+    lm_nostrils = lm[31: 36]  # top-down
+    lm_eye_left = lm[36: 42]  # left-clockwise
+    lm_eye_right = lm[42: 48]  # left-clockwise
+    lm_mouth_outer = lm[48: 60]  # left-clockwise
+    lm_mouth_inner = lm[60: 68]  # left-clockwise
+
+    # Calculate auxiliary vectors.
+    eye_left = np.mean(lm_eye_left, axis=0)
+    eye_right = np.mean(lm_eye_right, axis=0)
+    eye_avg = (eye_left + eye_right) * 0.5
+    eye_to_eye = eye_right - eye_left
+    mouth_left = lm_mouth_outer[0]
+    mouth_right = lm_mouth_outer[6]
+    mouth_avg = (mouth_left + mouth_right) * 0.5
+    eye_to_mouth = mouth_avg - eye_avg
+
+    # Choose oriented crop rectangle.
+    x = eye_to_eye - np.flipud(eye_to_mouth) * [-1, 1]
+    x /= np.hypot(*x)
+    x *= max(np.hypot(*eye_to_eye) * 2.0, np.hypot(*eye_to_mouth) * 1.8)
+    y = np.flipud(x) * [-1, 1]
+    c = eye_avg + eye_to_mouth * 0.1
+    quad = np.stack([c - x - y, c - x + y, c + x + y, c + x - y])
+    qsize = np.hypot(*x) * 2
+
+    # read image
+    transform_size = output_size
+    enable_padding = True
+
+    # Shrink.
+    shrink = int(np.floor(qsize / output_size * 0.5))
+    if shrink > 1:
+        rsize = (int(np.rint(float(img.size[0]) / shrink)), int(np.rint(float(img.size[1]) / shrink)))
+        img = img.resize(rsize, Image.ANTIALIAS)
+        quad /= shrink
+        qsize /= shrink
+
+    # Crop.
+    border = max(int(np.rint(qsize * 0.1)), 3)
+    crop = (int(np.floor(min(quad[:, 0]))), int(np.floor(min(quad[:, 1]))), int(np.ceil(max(quad[:, 0]))),
+            int(np.ceil(max(quad[:, 1]))))
+    crop = (max(crop[0] - border, 0), max(crop[1] - border, 0), min(crop[2] + border, img.size[0]),
+            min(crop[3] + border, img.size[1]))
+    if crop[2] - crop[0] < img.size[0] or crop[3] - crop[1] < img.size[1]:
+        img = img.crop(crop)
+        quad -= crop[0:2]
+
+    # Pad.
+    pad = (int(np.floor(min(quad[:, 0]))), int(np.floor(min(quad[:, 1]))), int(np.ceil(max(quad[:, 0]))),
+           int(np.ceil(max(quad[:, 1]))))
+    pad = (max(-pad[0] + border, 0), max(-pad[1] + border, 0), max(pad[2] - img.size[0] + border, 0),
+           max(pad[3] - img.size[1] + border, 0))
+    if enable_padding and max(pad) > border - 4:
+        pad = np.maximum(pad, int(np.rint(qsize * 0.3)))
+        img = np.pad(np.float32(img), ((pad[1], pad[3]), (pad[0], pad[2]), (0, 0)), 'reflect')
+        h, w, _ = img.shape
+        y, x, _ = np.ogrid[:h, :w, :1]
+        mask = np.maximum(1.0 - np.minimum(np.float32(x) / pad[0], np.float32(w - 1 - x) / pad[2]),
+                          1.0 - np.minimum(np.float32(y) / pad[1], np.float32(h - 1 - y) / pad[3]))
+        blur = qsize * 0.02
+        img += (scipy.ndimage.gaussian_filter(img, [blur, blur, 0]) - img) * np.clip(mask * 3.0 + 1.0, 0.0, 1.0)
+        img += (np.median(img, axis=(0, 1)) - img) * np.clip(mask, 0.0, 1.0)
+        img = Image.fromarray(np.uint8(np.clip(np.rint(img), 0, 255)), 'RGB')
+        quad += pad[:2]
+
+    # Transform.
+    img = img.transform((transform_size, transform_size), Image.QUAD, (quad + 0.5).flatten(), Image.BILINEAR)
+    if output_size < transform_size:
+        img = img.resize((output_size, output_size), Image.ANTIALIAS)
+
+    # Return aligned image.
+    return img