def encode_filter(filter_files):
images = []
faces = []
FACE_ALIGNMENT = FaceAlignment(LandmarksType._2D,
enable_cuda=True,
flip_input=False)
for i, filter_file in enumerate(filter_files):
images.append(skimage.io.imread(str(filter_file)))
faces.append(FACE_ALIGNMENT.get_landmarks(images[i]))
FACE_ALIGNMENT = None
face_recognition_model = face_recognition_models.face_recognition_model_location(
)
face_encoder = dlib.face_recognition_model_v1(face_recognition_model)
for i, face in enumerate(faces):
if face is None:
print('Warning: {} has no face.'.format(filter_files[i]))
continue
if len(face) > 1:
print('Warning: {} has more than one face.'.format(
filter_files[i]))
parts = []
for p in face[0]:
parts.append(dlib.point(p[0], p[1]))
raw_landmark_set = dlib.full_object_detection(rect, parts)
yield numpy.array(
face_encoder.compute_face_descriptor(images[i],
raw_landmark_set, 1))
def landmark_preprocess(noisy_path, dataset_path):
noisy_train_path = os.path.join(noisy_path, "train")
noisy_test_path = os.path.join(noisy_path, "test")
noisy_val_path = os.path.join(noisy_path, "validation")
train_path = os.path.join(dataset_path, "train")
test_path = os.path.join(dataset_path, "test")
val_path = os.path.join(dataset_path, "validation")
checkpath(dataset_path)
checkpath(train_path)
checkpath(test_path)
checkpath(val_path)
fa = FaceAlignment(LandmarksType._2D, device="cuda:1")
train_files = file_list(noisy_train_path, ".png")
for i in train_files:
landmarks = fa.get_landmarks_from_image(i)[0]
img = plot_landmarks((300, 300, 3), landmarks)
img.save(os.path.join(train_path, "lm" + i.split("y")[-1]))
test_files = file_list(noisy_test_path, ".png")
for i in test_files:
landmarks = fa.get_landmarks_from_image(i)[0]
img = plot_landmarks((300, 300, 3), landmarks)
img.save(os.path.join(test_path, "lm" + i.split("y")[-1]))
val_files = file_list(noisy_val_path, ".png")
print(noisy_val_path)
for i in val_files:
landmarks = fa.get_landmarks_from_image(i)[0]
img = plot_landmarks((300, 300, 3), landmarks)
img.save(os.path.join(val_path, "lm" + i.split("y")[-1]))
def __init__(self, style_img, input_img, style_mask, input_mask, save=False):
style_name = os.path.basename(style_img).split('.')[0]
input_name = os.path.basename(input_img).split('.')[0]
self.style_img = np.float32(imread(style_img))
self.input_img = np.float32(imread(input_img))
self.style_mask = np.float32(imread(style_mask))
self.input_mask = np.float32(imread(input_mask))
# Fetch Facial Landmarks
if os.path.exists('input/%s_%s_lm.pkl' % (style_name, input_name)):
with open('input/%s_%s_lm.pkl' % (style_name, input_name), 'rb') as f:
pkl = pickle.load(f)
self.style_lm = pkl['style']
self.input_lm = pkl['input']
else:
fa = FaceAlignment(LandmarksType._2D, device='cpu', flip_input=False)
self.style_lm = fa.get_landmarks(self.style_img)[0]
self.input_lm = fa.get_landmarks(self.input_img)[0]
with open('input/%s_%s_lm.pkl' % (style_name, input_name),
'wb') as f:
pickle.dump({
'style': self.style_lm,
'input': self.input_lm
}, f, protocol=2)
self.output_filename = '_'.join({input_name, style_name})
self.save = save
def compare(root, f1, f2):
global face_det
global face_recon
global face_align
if not face_det:
face_det = FaceDetection(gpu_id)
if not face_recon:
face_recon = FaceRecogniton(gpu_id)
if not face_align:
face_align = FaceAlignment(gpu_id)
time_start = time.time()
img_a = cv2.imread(root + '/' + f1)
img_b = cv2.imread(root + '/' + f2)
bbox_list1, a_point = face_det.get_max_bounding_box_by_image(img_a)
bbox_list2, b_point = face_det.get_max_bounding_box_by_image(img_b)
similarity = 0
if bbox_list1 and bbox_list2:
a_aligned_faces = face_align.affine_face(img_a, a_point)
b_aligned_faces = face_align.affine_face(img_b, b_point)
similarity = face_recon.face_compare(a_aligned_faces, b_aligned_faces)
#print similarity
time_end = time.time()
time_use = int(1000 * (time_end - time_start))
#print 'time_used:' + str(time_use)
return similarity, time_use
def __getitem__(self, idx):
real_idx = self.indexes[idx]
path = self.files[real_idx]
print("image file path=", path)
fa = FaceAlignment(LandmarksType._2D, device=self.device)
imgUMat = cv2.imread(path)
x_temp = cv2.cvtColor(imgUMat, cv2.COLOR_BGR2RGB)
y_temp = fa.get_landmarks(x_temp)[0]
out = []
x = PIL.Image.fromarray(x_temp, 'RGB')
y = plot_landmarks(x_temp, y_temp)
if self.transform:
x = self.transform(x)
y = self.transform(y)
out.append({'frame': x, 'landmarks': y})
return real_idx, out
def evaluate(respth='./results/data_src', dspth='../data'):
respth = osp.join(os.path.abspath(os.path.dirname(__file__)), respth)
if not os.path.exists(respth):
os.makedirs(respth)
face_model = FaceAlignment(LandmarksType._2D, device="cuda")
data_path = osp.join(os.path.abspath(os.path.dirname(__file__)), dspth)
for image_path in os.listdir(data_path):
image = cv2.imread(osp.join(data_path, image_path))
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
landmark = face_model.get_landmarks(image)[-1]
# print(landmark)
mask = get_image_hull_mask(np.shape(image), landmark).astype(np.uint8)
# cv2.imshow("mask", (mask*255).astype(np.uint8))
image_bgra = merge(image, mask)
# cv2.imshow("image_bgra", image_bgra)
# cv2.waitKey(1)
save_path = osp.join(respth, image_path)
cv2.imwrite(save_path[:-4] + '.png', image_bgra)
def eval(input_path, output_path, checkpoint_path, model, gpu):
input = Image.open(input_path)
input = input.convert("RGB")
w, h = input.size
w_, h_ = 128 * (w // 128), 128 * (h // 128)
fa = FaceAlignment(LandmarksType._2D, device="cuda:" + str(gpu))
landmarks = fa.get_landmarks_from_image(input_path)[0]
landmark_img = plot_landmarks(np.array(input), landmarks)
transform_forward = transforms.Compose([
transforms.Resize((w_, h_)),
transforms.CenterCrop((w_, h_)),
transforms.ToTensor()
])
transform_backward = transforms.Compose([
transforms.ToPILImage(),
transforms.Resize((w, h)),
transforms.CenterCrop((w, h)),
])
input = transform_forward(input)
landmark_img = transform_forward(landmark_img)
if model == "Pix2Pix":
NFNet = Pix2Pix()
else:
NFNet = ResResNet()
checkpoint = torch.load(checkpoint_path)
NFNet.load_state_dict(checkpoint['my_classifier'])
NFNet.to(gpu)
x = torch.cat((input, landmark_img), 0)
x = x.unsqueeze(0)
x = x.to(gpu)
output = NFNet(x)
output = output.to("cpu")
output = transform_backward(output[0])
output.save(output_path)
def preprocess_dataset(source,
output,
device='cpu',
size=0,
overwrite=False,
frame_rate=1):
"""
Starts the pre-processing of the VoxCeleb dataset used for the Talking Heads models. This process has the following
steps:
* Extract all frames of each video in the dataset. Frames of videos that are split in several files are joined
together.
* Select K+1 frames of each video that will be kept. K frames will be used to train the embedder network, while
the other one will be used to train the generator network. The value of K can be configured in the config.py file.
* Landmarks will be extracted for the face in each of the frames that are being kept.
* The frames and the corresponding landmarks for each video will be saved in files (one for each video) in the
output directory.
We originally tried to process several videos simultaneously using multiprocessing, but this seems to actually
slow down the process instead of speeding it up.
:param source: Path to the raw VoxCeleb dataset.
:param output: Path where the pre-processed videos will be stored.
:param device: Device used to run the landmark extraction model.
:param size: Size of the dataset to generate. If 0, the entire raw dataset will be processed, otherwise, as many
videos will be processed as specified by this parameter.
:param overwrite: f True, files that have already been processed will be overwritten, otherwise, they will be
ignored and instead, different files will be loaded.
"""
logging.info('===== DATASET PRE-PROCESSING =====')
logging.info(f'Running on {device.upper()}.')
logging.info(f'Saving K+1 random frames from each video (K = {config.K}).')
fa = FaceAlignment(LandmarksType._2D, device=device)
video_list = get_video_list(source, size, output, overwrite=overwrite)
logging.info(f'Processing {len(video_list)} videos...')
# pool = Pool(processes=4, initializer=init_pool, initargs=(fa, output))
# pool.map(process_video_folder, video_list)
init_pool(fa, output)
counter = 1
for v in video_list:
start_time = datetime.now()
process_video_folder(v, frame_rate)
logging.info(
f'{counter}/{len(video_list)}\t{datetime.now()-start_time}')
counter += 1
logging.info(f'All {len(video_list)} videos processed.')
def __init__(self, args):
self.args = args
model = edict()
self.threshold = args.threshold
self.det_minsize = 50
self.det_threshold = [0.4, 0.6, 0.6]
self.det_factor = 0.9
_vec = args.image_size.split(',')
assert len(_vec) == 2
image_size = (int(_vec[0]), int(_vec[1]))
self.image_size = image_size
_vec = args.model.split(',')
assert len(_vec) == 2
prefix = _vec[0]
epoch = int(_vec[1])
print('loading', prefix, epoch)
ctx = mx.gpu(args.gpu)
sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, epoch)
all_layers = sym.get_internals()
sym = all_layers['fc1_output']
model = mx.mod.Module(symbol=sym, context=ctx, label_names=None)
#model.bind(data_shapes=[('data', (args.batch_size, 3, image_size[0], image_size[1]))], label_shapes=[('softmax_label', (args.batch_size,))])
model.bind(data_shapes=[('data', (1, 3, image_size[0],
image_size[1]))])
model.set_params(arg_params, aux_params)
self.model = model
# mtcnn_path = os.path.join(os.path.dirname(__file__), 'mtcnn-model')
mtcnn_path = os.path.join('deploy', 'mtcnn-model')
detector = MtcnnDetector(model_folder=mtcnn_path,
ctx=ctx,
num_worker=1,
accurate_landmark=True,
threshold=[0.0, 0.0, 0.2])
self.detector = detector
self.FACE_ALIGNMENT = FaceAlignment(LandmarksType._3D,
device='cuda',
flip_input=False)
def preprocess_dataset(source, output, device='cpu', size=0, overwrite=False):
logging.info('===== DATASET PRE-PROCESSING =====')
logging.info(f'Running on {device.upper()}.')
logging.info(f'Saving K+1 random frames from each video (K = {K}).')
fa = FaceAlignment(LandmarksType._2D, device=device)
video_list = get_video_list(source, size, output, overwrite=overwrite)
logging.info(f'Processing {len(video_list)} videos...')
init_pool(fa, output)
counter = 1
for v in video_list:
process_video_folder(v)
logging.info(f'{counter}/{len(video_list)}')
counter += 1
logging.info(f'All {len(video_list)} videos processed.')
def getTransform(videoName, modelIdx):
modelList = [
'sw', 'han', 'tsai', 'father', 'cloud', 'aerith', 'tifa', 'davinci'
]
G = network.Generator()
G = load_model(G,
"app/modules/talkingHeads/resource/" + modelList[modelIdx],
modelList[modelIdx])
G = G.to("cuda:0")
fa = FaceAlignment(LandmarksType._2D, device='cuda:0')
e_vector = get_e_vector("app/modules/talkingHeads/resource/" +
modelList[modelIdx] + "/" + modelList[modelIdx] +
".npy")
timestamp = str(int(time.time()))
print(timestamp)
# generate_moving_video(G, "app/static/"+videoName, "app/modules/talkingHeads/resource/"+modelList[modelIdx]+"/"+modelList[modelIdx]+".npy", "app/static/result-"+timestamp+".mp4", "cuda:0")
generate_moving_video(G, "app/static/" + videoName, e_vector,
"app/static/result-" + timestamp + ".mp4", "cuda:0",
fa)
return jsonify({"code": 200, "message": "轉換成功", 'token': timestamp})
def end2end_mask_encoding(
image: np.ndarray,
face_aligment_class: face_alignment.FaceAlignment) -> dict:
landmarks = face_aligment_class.get_landmarks_from_image(image)
landmarks = np.floor(landmarks[0]).astype(np.int32)
target_points, s1, s2 = extract_target_points_and_characteristic(landmarks)
mask_rgba_crop, target_points = extract_polygon(image, target_points)
mask_rgba_crop, target_points = rotate_image_and_points(
mask_rgba_crop, s1, target_points)
res = {
's1': s1,
's2': s2,
'points': target_points.tolist(),
'base64_img': image_to_string(mask_rgba_crop)
}
return res
def end2end_mask_generation(image: np.ndarray,
masks_database: dict,
face_aligment_class: face_alignment.FaceAlignment,
input_face_landmarks: np.ndarray = None):
if input_face_landmarks is None:
face_landmarks = face_aligment_class.get_landmarks_from_image(image)
if len(face_landmarks) == 0:
raise RuntimeError('Can\'t find facial landmarks')
face_landmarks = np.floor(face_landmarks[0]).astype(np.int32)
else:
face_landmarks = input_face_landmarks
_, _, s2 = extract_target_points_and_characteristic(face_landmarks)
sampling_mask_data = extract_mask_from_base(masks_database, s2)
face_with_mask = apply_mask_to_image_with_face(image, sampling_mask_data,
face_landmarks)
return face_with_mask
def monkey_patch_face_detector(_):
detector = dlib.get_frontal_face_detector()
class Rect(object):
def __init__(self, rect):
self.rect = rect
def detect(*args):
return [Rect(x) for x in detector(*args)]
return detect
dlib.cnn_face_detection_model_v1 = monkey_patch_face_detector
FACE_ALIGNMENT = FaceAlignment(LandmarksType._2D,
enable_cuda=True,
flip_input=False)
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,
e_vector = get_e_vector("app/modules/talkingHeads/resource/" +
modelList[modelIdx] + "/" + modelList[modelIdx] +
".npy")
timestamp = str(int(time.time()))
print(timestamp)
# generate_moving_video(G, "app/static/"+videoName, "app/modules/talkingHeads/resource/"+modelList[modelIdx]+"/"+modelList[modelIdx]+".npy", "app/static/result-"+timestamp+".mp4", "cuda:0")
generate_moving_video(G, "app/static/" + videoName, e_vector,
"app/static/result-" + timestamp + ".mp4", "cuda:0",
fa)
return jsonify({"code": 200, "message": "轉換成功", 'token': timestamp})
G = network.Generator()
G = load_model(G, "app/modules/talkingHeads/resource/han", "han")
G = G.to("cuda:0")
fa = FaceAlignment(LandmarksType._2D, device='cuda:0')
e_vector = get_e_vector("app/modules/talkingHeads/resource/han/han.npy")
def imgTransform(srcImage, modelIdx):
image = base64_cv2(srcImage)
# image = cv2.imread("app/modules/2.png")
image = cv2.resize(image, (256, 256))
# modelList=['sw','han','tsai']
# G = network.Generator()
# G = load_model(G, "app/modules/talkingHeads/resource/"+modelList[modelIdx], modelList[modelIdx])
# G = G.to("cuda:0")
# fa = FaceAlignment(LandmarksType._2D, device='cuda:0')
# e_vector = get_e_vector("app/modules/talkingHeads/resource/"+modelList[modelIdx]+"/"+modelList[modelIdx]+".npy")
result = generate_moving_image(G, image, e_vector, "cuda:0", fa)
print('don')
from PIL import Image
import regex
import torch
import subprocess
import hashlib
import sys
from demo import load_checkpoints
from animate import normalize_kp
app = Flask(__name__)
generator, kp_detector = load_checkpoints(
config_path="first-order-model/config/vox-adv-256.yaml",
checkpoint_path="vox-adv-cpk.pth.tar",
)
fa = FaceAlignment(LandmarksType._2D)
@app.route("/")
def index():
return render_template("upload.html")
def data(obj) -> str:
return f"data: {json.dumps(obj)}\n\n"
@app.route("/upload", methods=["POST"])
def upload():
for key, file in request.files.items():
pathlib.Path("static", key).mkdir(exist_ok=True)
import cv2
import random
import numpy as np
import pickle as pkl
import tensorflow as tf
from tqdm import tqdm
import pandas as pd
from functools import partial
import multiprocessing as mp
from imutils import face_utils
from hyperparams import Hyperparams as hp
from face_alignment import FaceAlignment, LandmarksType
from utils import detector, predictor, preprocess_input
global face_alignment
face_alignment = FaceAlignment(LandmarksType._2D, device='cuda')
def get_video_list(source = hp.dataset):
"""
Extracts a list of paths to videos to pre-process during the current run.
:param source: Path to the root directory of the dataset.
:return: List of paths to videos.
"""
video_list = []
for root, dirs, files in tqdm(os.walk(source)):
if len(files) > 0:
assert contains_only_videos(files) and len(dirs) == 0
video_list.append((root, files))
def __init__(self, dimensions='2d'):
landmarkType = LandmarksType._2D if dimensions == '2d' else LandmarksType._3D
self.faceAlignment = FaceAlignment(landmarkType,
flip_input=False,
device='cpu',
verbose=False)
class FaceTools:
""" Face Toolkit
"""
def __init__(self, dimensions='2d'):
landmarkType = LandmarksType._2D if dimensions == '2d' else LandmarksType._3D
self.faceAlignment = FaceAlignment(landmarkType,
flip_input=False,
device='cpu',
verbose=False)
@staticmethod
def _polyArea(points):
""" calculate the area of a polygon (Gauss equation)
"""
x = [point[0] for point in points]
y = [point[1] for point in points]
return 0.5 * np.abs(
np.dot(x, np.roll(y, 1)) - np.dot(y, np.roll(x, 1)))
@staticmethod
def _slope(point1, point2):
""" compute the slope of the line passing through two points
"""
return (point2[1] - point1[1]) / (point2[0] - point1[0])
def landmarks(self, img):
""" get landmarks
"""
return self.faceAlignment.get_landmarks(img)[0]
def geometricFeatures1(self, img, landmarks=None):
"""Improved Performance in Facial Expression Recognition Using 32 Geometric Features
Linear features(15):
– 3 for left eyebrow
– 2 for left eye
– 1 for cheeks
– 1 for nose
– 8 for mouth
Polygonal features(3):
– 1 for the left eye
– 1 between corners of left eye and left corner of mouth
– 1 for mouth
Elliptical features(7):
– 1 for left eyebrow
– 3 for the left eye: eye, upper and lower eyelids
– 3 for mouth: upper and lower lips
Slope features(7):
– 1 for left eyebrow
– 6 for mouth corners
"""
pts = self.faceAlignment.get_landmarks(
img)[0] if landmarks is None else landmarks
result = []
# Linear features
#eyebrow
result.append(d(pts[21], pts[22]))
result.append(d(pts[22], pts[42]))
result.append(d(pts[26], pts[45]))
#left eye
result.append(d(pts[42], pts[45]))
#d2=43-47 || 44-46
result.append(d(pts[43], pts[47]))
#cheeks
result.append(d(pts[1], pts[15]))
#nose
result.append(d(pts[31], pts[35]))
#mouth
result.append(d(pts[48], pts[51]))
result.append(d(pts[51], pts[54]))
result.append(d(pts[62], pts[66]))
result.append(d(pts[51], pts[57]))
result.append(d(pts[50], pts[33]))
result.append(d(pts[52], pts[33]))
result.append(d(pts[48], pts[31]))
result.append(d(pts[54], pts[35]))
#Polygonal features:
result.append(FaceTools._polyArea([pts[48], pts[54], pts[57]]))
result.append(FaceTools._polyArea([pts[54], pts[42], pts[45]]))
result.append(FaceTools._polyArea([pts[42], pts[22], pts[26],
pts[45]]))
#Slope features:
result.append(FaceTools._slope(pts[22], pts[26]))
result.append(FaceTools._slope(pts[48], pts[31]))
result.append(FaceTools._slope(pts[54], pts[35]))
result.append(FaceTools._slope(pts[48], pts[51]))
result.append(FaceTools._slope(pts[51], pts[54]))
result.append(FaceTools._slope(pts[54], pts[57]))
result.append(FaceTools._slope(pts[48], pts[57]))
return result
def faceDistances(self, img, landmarks=None):
""" calculate distances as described in
'Automatic Facial Expression Recognition Using Combined Geometric Features':
D1 Left eyebrow length
D2 Right eyebrow length
D3 Distance between left and right eyebrow
D4 Left eye height
D5 Left eye width
D6 Right eye height
D7 Right eye width
D8 Distance between left eyebrow and left eye
D9 Distance between right eyebrow and right eye
D10 Distance between nose tip and upper lip
D11 Lip width
D12 Lip height
D13 Inner lip distance
D14 Distance between left eye corner and lip left corner
D15 Distance between right eye corner and lip right corner
"""
pts = self.faceAlignment.get_landmarks(
img)[0] if landmarks is None else landmarks
distances = [
d(pts[17], pts[18]) + d(pts[18], pts[19]) + d(pts[19], pts[20]) + \
d(pts[20], pts[21]),
d(pts[22], pts[23]) + d(pts[23], pts[24]) + d(pts[24], pts[25]) + \
d(pts[25], pts[26]),
d(pts[21], pts[22]),
d(pts[40], pts[38]),
d(pts[36], pts[39]),
d(pts[43], pts[47]),
d(pts[42], pts[45]),
d(pts[19], pts[37]),
d(pts[24], pts[44]),
d(pts[33], pts[51]),
d(pts[48], pts[54]),
d(pts[51], pts[57]),
d(pts[62], pts[66]),
d(pts[36], pts[48]),
d(pts[45], pts[54])
]
return distances
class FaceModel:
def __init__(self, args):
self.args = args
model = edict()
self.threshold = args.threshold
self.det_minsize = 50
self.det_threshold = [0.4, 0.6, 0.6]
self.det_factor = 0.9
_vec = args.image_size.split(',')
assert len(_vec) == 2
image_size = (int(_vec[0]), int(_vec[1]))
self.image_size = image_size
_vec = args.model.split(',')
assert len(_vec) == 2
prefix = _vec[0]
epoch = int(_vec[1])
print('loading', prefix, epoch)
ctx = mx.gpu(args.gpu)
sym, arg_params, aux_params = mx.model.load_checkpoint(prefix, epoch)
all_layers = sym.get_internals()
sym = all_layers['fc1_output']
model = mx.mod.Module(symbol=sym, context=ctx, label_names=None)
#model.bind(data_shapes=[('data', (args.batch_size, 3, image_size[0], image_size[1]))], label_shapes=[('softmax_label', (args.batch_size,))])
model.bind(data_shapes=[('data', (1, 3, image_size[0],
image_size[1]))])
model.set_params(arg_params, aux_params)
self.model = model
# mtcnn_path = os.path.join(os.path.dirname(__file__), 'mtcnn-model')
mtcnn_path = os.path.join('deploy', 'mtcnn-model')
detector = MtcnnDetector(model_folder=mtcnn_path,
ctx=ctx,
num_worker=1,
accurate_landmark=True,
threshold=[0.0, 0.0, 0.2])
self.detector = detector
self.FACE_ALIGNMENT = FaceAlignment(LandmarksType._3D,
device='cuda',
flip_input=False)
def get_feature(self, face_img):
detected = True
#face_img is bgr image
def mtcnn_align(img):
ret = self.detector.detect_face_limited(face_img,
det_type=self.args.det)
if ret is None:
# detected = False
bbox, points = None, None
else:
bbox, points = ret
if bbox.shape[0] == 0:
# detected = False
bbox, points = None, None
else:
bbox = bbox[0, 0:4]
points = points[0, :].reshape((2, 5)).T
# print(bbox)
# print(points)
nimg = face_preprocess.preprocess(face_img,
bbox,
points,
image_size='112,112')
return nimg
# skimage.io.imread( str(fn) )
faces = self.FACE_ALIGNMENT.get_landmarks(face_img)
if faces is not None:
if len(faces) > 1:
faces = faces[0:1]
# pdb.set_trace()
points = faces[0]
alignment = umeyama(points[17:], landmarks_2D, True)[0:2]
nimg = _aligned_image = transform(face_img, alignment, 112, 0)
else:
nimg = mtcnn_align(face_img)
# nimg = face_preprocess.preprocess(face_img, image_size='112,112')
# pdb.set_trace()
nimg = cv2.cvtColor(nimg, cv2.COLOR_BGR2RGB)
aligned = np.transpose(nimg, (2, 0, 1))
#print(nimg.shape)
embedding = None
for flipid in [0, 1]:
if flipid == 1:
if self.args.flip == 0:
break
do_flip(aligned)
input_blob = np.expand_dims(aligned, axis=0)
data = mx.nd.array(input_blob)
db = mx.io.DataBatch(data=(data, ))
self.model.forward(db, is_train=False)
_embedding = self.model.get_outputs()[0].asnumpy()
#print(_embedding.shape)
if embedding is None:
embedding = _embedding
else:
embedding += _embedding
embedding = sklearn.preprocessing.normalize(embedding).flatten()
return detected, embedding