def video(video_path_or_cam):
facer = FaceAna()
vide_capture = cv2.VideoCapture(video_path_or_cam)
while 1:
ret, image = vide_capture.read()
pattern = np.zeros_like(image)
img_show = image.copy()
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
star = time.time()
boxes, landmarks, states = facer.run(image)
duration = time.time() - star
print('one iamge cost %f s' % (duration))
for face_index in range(landmarks.shape[0]):
#######head pose need develop
# reprojectdst, euler_angle=get_head_pose(landmarks[face_index],img_show)
if args.mask:
face_bbox_keypoints = np.concatenate(
(landmarks[face_index][:17, :],
np.flip(landmarks[face_index][17:27, :], axis=0)),
axis=0)
pattern = cv2.fillPoly(pattern,
[face_bbox_keypoints.astype(np.int)],
(1., 1., 1.))
# for start, end in line_pairs:
# cv2.line(img_show, reprojectdst[start], reprojectdst[end], (0, 0, 255),2)
#
# cv2.putText(img_show, "X: " + "{:7.2f}".format(euler_angle[0, 0]), (20, 20), cv2.FONT_HERSHEY_SIMPLEX,
# 0.75, (0, 0, 0), thickness=2)
# cv2.putText(img_show, "Y: " + "{:7.2f}".format(euler_angle[1, 0]), (20, 50), cv2.FONT_HERSHEY_SIMPLEX,
# 0.75, (0, 0, 0), thickness=2)
# cv2.putText(img_show, "Z: " + "{:7.2f}".format(euler_angle[2, 0]), (20, 80), cv2.FONT_HERSHEY_SIMPLEX,
# 0.75, (0, 0, 0), thickness=2)
for landmarks_index in range(landmarks[face_index].shape[0]):
x_y = landmarks[face_index][landmarks_index]
cv2.circle(img_show, (int(x_y[0]), int(x_y[1])), 3,
(222, 222, 222), -1)
cv2.namedWindow("capture", 0)
cv2.imshow("capture", img_show)
if args.mask:
cv2.namedWindow("masked", 0)
cv2.imshow("masked", image * pattern)
key = cv2.waitKey(1)
if key == ord('q'):
return
def images(image_dir):
facer = FaceAna()
image_list = os.listdir(image_dir)
image_list = [x for x in image_list if 'jpg' in x or 'png' in x]
image_list.sort()
for idx, image_name in enumerate(image_list):
image = cv2.imread(os.path.join(image_dir, image_name))
pattern = np.zeros_like(image)
img_show = image.copy()
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
star = time.time()
boxes, landmarks, states = facer.run(image)
###no track
facer.reset()
duration = time.time() - star
print('%d %s time: %fs' % (idx, image_name, duration))
for face_index in range(landmarks.shape[0]):
#######head pose
#reprojectdst, euler_angle=get_head_pose(landmarks[face_index],img_show)
if args.mask:
face_bbox_keypoints = np.concatenate(
(landmarks[face_index][:17, :],
np.flip(landmarks[face_index][17:27, :], axis=0)),
axis=0)
pattern = cv2.fillPoly(pattern,
[face_bbox_keypoints.astype(np.int)],
(1., 1., 1.))
# for start, end in line_pairs:
# cv2.line(img_show, reprojectdst[start], reprojectdst[end], (0, 0, 255),2)
#
# cv2.putText(img_show, "X: " + "{:7.2f}".format(euler_angle[0, 0]), (20, 20), cv2.FONT_HERSHEY_SIMPLEX,
# 0.75, (0, 0, 0), thickness=2)
# cv2.putText(img_show, "Y: " + "{:7.2f}".format(euler_angle[1, 0]), (20, 50), cv2.FONT_HERSHEY_SIMPLEX,
# 0.75, (0, 0, 0), thickness=2)
# cv2.putText(img_show, "Z: " + "{:7.2f}".format(euler_angle[2, 0]), (20, 80), cv2.FONT_HERSHEY_SIMPLEX,
# 0.75, (0, 0, 0), thickness=2)
for landmarks_index in range(landmarks[face_index].shape[0]):
x_y = landmarks[face_index][landmarks_index]
cv2.circle(img_show, (int(x_y[0]), int(x_y[1])), 3,
(222, 222, 222), -1)
cv2.imwrite("results/" + image_name, img_show)
import cv2
import time
import imageio
import numpy as np
from lib.core.api.facer import FaceAna
from lib.core.headpose.pose import get_head_pose, line_pairs
facer = FaceAna()
def video(video_path_or_cam):
start_holder = False
vide_capture = cv2.VideoCapture(video_path_or_cam)
buff = []
counter = 1
while 1:
counter += 1
ret, image = vide_capture.read()
#image = image[100:1000, :, :]
image = cv2.resize(image, None, fx=2., fy=2.)
pattern = np.zeros_like(image)
img_show = image.copy()
img = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
star = time.time()
boxes, landmarks, states = facer.run(img)
import wear_mask
from PIL import Image
import numpy as np
import cv2
import os
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
from lib.core.api.facer import FaceAna
# image = cv2.imread('test_img/01.jpg')
image = cv2.imread('test_img/Aishwarya_Rai_0001.png')
facer = FaceAna()
boxes, landmarks, _ = facer.run(image)
face_image = wear_mask.mask_img('test_img/Aishwarya_Rai_0001.png', 'images',
landmarks)
# cv2.imshow(face_image)
# face_image = np.array(face_image)
face_image.show()
from lib.core.api.facer import FaceAna
import numpy as np
import cv2
def to_np_array(image_data, dtype=np.uint8):
width = image_data.width
height = image_data.height
pixel_length = image_data.pixel_length
data = image_data.data
assert len(data) == width * height * pixel_length
return np.frombuffer(data, dtype=dtype).reshape(
(height, width, pixel_length))
facer = FaceAna()
# remind 消除初始化时间
boxes, landmarks, states = facer.run(
cv2.imread("D:\pycharmProject\Peppa_Pig_Face_Engine\img\\test_32.jpg"))
class GrpcServer(service.FaceServerServicer):
def predict(self, request, context):
star = time.time()
# print(request.frame)
image = to_np_array(request.img)
# print(image.shape)
# cv2.imwrite("D:\pycharmProject\Peppa_Pig_Face_Engine\img\\test_%d.jpg" % (request.frame), image)
# print(request.frame, 'reshape cost %f s' % (time.time() - star))
boxes, landmarks, states = facer.run(image)
print(request.frame, 'detect cost %f s' % (time.time() - star))
def __init__(self):
self.facer = FaceAna()
self.frame = 0
self.inc = 1
class Test:
def __init__(self):
self.facer = FaceAna()
self.frame = 0
self.inc = 1
# 被周期性调度触发函数
def printTime(self):
print(datetime.now().strftime("%Y-%m-%d %H:%M:%S"), self.frame)
self.frame = 0
t = Timer(self.inc, self.printTime, ())
t.start()
def video(self, video_path_or_cam):
vide_capture = cv2.VideoCapture(video_path_or_cam)
# vide_capture.set(3, 640)
# vide_capture.set(4, 480)
while 1:
ret, image = vide_capture.read()
self.frame += 1
pattern = np.zeros_like(image)
img_show = image.copy()
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
star = time.time()
boxes, landmarks, states = self.facer.run(image)
duration = time.time() - star
# print('one iamge cost %f s' % (duration))
for face_index in range(landmarks.shape[0]):
#######head pose
reprojectdst, euler_angle = get_head_pose(
landmarks[face_index], img_show)
if args.mask:
face_bbox_keypoints = np.concatenate(
(landmarks[face_index][:17, :],
np.flip(landmarks[face_index][17:27, :], axis=0)),
axis=0)
pattern = cv2.fillPoly(
pattern, [face_bbox_keypoints.astype(np.int)],
(1., 1., 1.))
for start, end in line_pairs:
cv2.line(img_show, reprojectdst[start], reprojectdst[end],
(0, 0, 255), 2)
cv2.putText(img_show,
"X: " + "{:7.2f}".format(euler_angle[0, 0]),
(20, 20),
cv2.FONT_HERSHEY_SIMPLEX,
0.75, (0, 0, 0),
thickness=2)
cv2.putText(img_show,
"Y: " + "{:7.2f}".format(euler_angle[1, 0]),
(20, 50),
cv2.FONT_HERSHEY_SIMPLEX,
0.75, (0, 0, 0),
thickness=2)
cv2.putText(img_show,
"Z: " + "{:7.2f}".format(euler_angle[2, 0]),
(20, 80),
cv2.FONT_HERSHEY_SIMPLEX,
0.75, (0, 0, 0),
thickness=2)
for landmarks_index in range(landmarks[face_index].shape[0]):
x_y = landmarks[face_index][landmarks_index]
cv2.circle(img_show, (int(x_y[0]), int(x_y[1])), 3,
(222, 222, 222), -1)
cv2.namedWindow("capture", 0)
cv2.imshow("capture", img_show)
if args.mask:
cv2.namedWindow("masked", 0)
cv2.imshow("masked", image * pattern)
key = cv2.waitKey(1)
if key == ord('q'):
return
def processVideo(self):
if self.mode == 'train':
from lib.core.api.facer import FaceAna
from lib.core.headpose.pose import get_head_pose, line_pairs
print('--------------- begin to load model -----------------------')
facer = FaceAna()
print('---------------- load model succcess ----------------------')
print('process video_path ==> %s ' % self.video_path)
# fil = open(self.label_path, 'a+')
margin = 25
crop_H = 400
crop_W = 400
## open video
video_capture=cv2.VideoCapture(self.video_path)
fps = video_capture.get(5)
frames = int(video_capture.get(7))
fourcc = cv2.VideoWriter_fourcc(*'mp4v') #opencv3.0
videoWriter = cv2.VideoWriter(self.save_video_path, fourcc, fps, (crop_W, crop_H))
RGB_mean = np.zeros((frames, 18)) # RGB mean value of six ROIs
landmarksAll = np.zeros((frames, 68, 2))
for i in range(frames):
valid_frame = 0
ret, image = video_capture.read()
# print(image.shape)
img_show = image.copy()
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)## img_show is BGR format
if self.mode == 'train':
boxes, landmarks, states = facer.run(image) ## image is RGB format
else:
self.videoFPS = fps
self.heartRate = ''
with open(self.landmark_path, 'r') as fil:
lines = fil.readlines()
line = lines[i].strip().split()
line = [int(item) for item in line]
landmarks = np.zeros((1, 68, 2), dtype='int32')
landmarks[0, :, 0] = line[4::2]
landmarks[0, :, 1] = line[5::2]
# assert landmarks.shape[0]==1, 'More than one person or no person in this video'
crop_img = np.zeros((crop_H, crop_W, 3))
face_index = 0 ### 默认一张脸
left_eye = landmarks[face_index][36:42]
right_eye = landmarks[face_index][42:48]
nose = landmarks[face_index][27:36]
face = landmarks[face_index][0:27]
leftup_eye_h = int(np.min(left_eye[..., 1])) # 左眼睛高度
leftdown_eye_h = int(np.max(left_eye[..., 1])) # 左眼睛高度
leftleft_eye_w = int(np.min(left_eye[..., 0])) # 左眼睛左端
leftright_eye_w = int(np.max(left_eye[..., 0])) # 左眼睛右端
left_eye_box1 = (leftleft_eye_w, leftdown_eye_h+5)
left_eye_box2 = (leftright_eye_w, 2*leftdown_eye_h-leftup_eye_h+5)
rightup_eye_h = int(np.min(right_eye[..., 1]))
rightdown_eye_h = int(np.max(right_eye[..., 1]))
rightleft_eye_w = int(np.min(right_eye[..., 0]))
rightright_eye_w = int(np.max(right_eye[..., 0]))
right_eye_box1 = (rightleft_eye_w, rightdown_eye_h+5)
right_eye_box2 = (rightright_eye_w, 2*rightdown_eye_h-rightup_eye_h+5)
noseup_h = int(np.min(nose[..., 1]))
nosedown_h = int(np.max(nose[..., 1]))
noseleft_w = int(np.min(nose[..., 0]))
noseright_w = int(np.max(nose[..., 0]))
faceup_h = int(np.min(face[..., 1]))
facedown_h = min(image.shape[0], int(np.max(face[..., 1]))) ## 防止超出边界，下巴不在视频中也会被预测到
faceleft_w = int(np.min(face[..., 0]))
faceright_w = int(np.max(face[..., 0]))
#start_left = int((crop_W-faceright_w+faceleft_w)/2)
#start_up = int((crop_H - facedown_h + faceup_h)/2)
#crop_img[start_up:start_up+facedown_h-faceup_h, start_left:start_left+faceright_w-faceleft_w, :] = img_show[faceup_h:facedown_h, faceleft_w:faceright_w, :]
#videoWriter.write(crop_img.astype('uint8'))
window_H = facedown_h - faceup_h
window_W = faceright_w - faceleft_w ### window_H和window_W是截取的框的大小
if window_H>window_W:
## 截取高度大于宽度，扩大宽度
window_up_h = faceup_h
window_down_h = facedown_h
diff = int((window_H-window_W)/2)
window_left_w = faceleft_w - diff
window_right_w = window_left_w + window_H
window_W = window_H
elif window_H < window_W:
## 截取宽度大于高度，扩大高度
window_left_w = faceleft_w
window_right_w = faceright_w
diff = int((window_W-window_H)/2)
window_up_h = faceup_h - diff
window_down_h = window_up_h + window_W
window_H = window_W
else:
window_up_h = faceup_h
window_down_h =facedown_h
window_left_w = faceleft_w
window_right_w = faceright_w
### (window_up_h, window_left_w)和(window_down_h, window_right_w)构成截取框
factor = crop_H/window_H ## 缩放倍数
tmp_img = img_show[window_up_h:window_down_h, window_left_w:window_right_w, :].astype('uint8')
crop_img = cv2.resize(tmp_img, (crop_H, crop_W),
interpolation=cv2.INTER_CUBIC).astype('uint8') ## cubic插值缩放
# videoWriter.write(crop_img)
# plt.imshow(crop_img[..., ::-1])
# plt.show()
"""
处理数据
"""
rela_coordinates = np.zeros((1, 2)) # 由于crop引入的相对坐标
rela_coordinates[0, 0] = int((window_up_h+window_down_h-crop_H)/2)
rela_coordinates[0, 1] = int((window_left_w+window_right_w-crop_W)/2)
# landmarksAll[i, ...] = (landmarks[face_index, :, ::-1] - rela_coordinates)*factor ## 保存相对坐标(h, w)，landmarks为(w, h)
"""
extract region of interest
"""
roi_1 = img_show[noseup_h-10:nosedown_h,noseleft_w:noseright_w, ::-1] ## BGR->RGB
roi_2 = img_show[min(leftup_eye_h, rightup_eye_h):max(leftdown_eye_h, rightdown_eye_h), leftright_eye_w:rightleft_eye_w, ::-1] ## BGR->RGB
roi_3 = img_show[leftdown_eye_h+5:2*leftdown_eye_h-leftup_eye_h+5, leftleft_eye_w:leftright_eye_w, ::-1] ## BGR->RGB
roi_4 = img_show[rightdown_eye_h+5:2*rightdown_eye_h-rightup_eye_h+5, rightleft_eye_w:rightright_eye_w, ::-1] ## BGR->RGB
mask_6 = np.zeros(img_show.shape[0:2]) ## region 5的mask
mask_6[faceup_h:facedown_h, leftleft_eye_w:rightright_eye_w] = 1.0
mask_face = np.zeros(img_show.shape[0:2]) ## 脸的mask
land1 = landmarks[face_index][0:17, ...].reshape(1, 17, 2)
land2 = landmarks[face_index][26:16:-1, ...].reshape(1, 10, 2)
land = np.concatenate([land1, land2], axis=1).astype('int32') ## 逆时针的脸部landmarks
cv2.polylines(mask_face, land, 1, 255)
cv2.fillPoly(mask_face, land, 255)
mask_face = mask_face/255.0
roi_5 = img_show.transpose((2, 0, 1))[::-1, ...] ##BGR->RGB
roi_5 = roi_5*mask_face
roi_5 = roi_5.reshape((3, -1))
mask_6 = mask_6*mask_face
roi_6 = img_show.transpose((2, 0, 1))[::-1, ...] ## BGR->RGB
roi_6 = roi_6*mask_6
roi_6 = roi_6.reshape((3, -1))
"""
计算RGB平均值
"""
RGB_1 = np.mean(np.reshape(roi_1/255.0, (-1, 3)), axis=0)
RGB_2 = np.mean(np.reshape(roi_2/255.0, (-1, 3)), axis=0)
RGB_3 = np.mean(np.reshape(roi_3/255.0, (-1, 3)), axis=0)
RGB_4 = np.mean(np.reshape(roi_4/255.0, (-1, 3)), axis=0)
RGB_5 = np.sum(roi_5/255.0, axis=1)/np.sum(mask_face)
RGB_6 = np.sum(roi_6/255.0, axis=1)/np.sum(mask_6)
RGB = np.concatenate([RGB_1, RGB_2, RGB_3, RGB_4, RGB_5, RGB_6])
if np.sum(np.isnan(RGB)) == 0: ##valid value
landmarksAll[valid_frame, ...] = (landmarks[face_index, :, ::-1] - rela_coordinates)*factor ## 保存相对坐标(h, w)，landmarks为(w, h)
RGB_mean[valid_frame, ...] = RGB ## save valid RGB
videoWriter.write(crop_img)
valid_frame += 1
# videoWriter.release() ## save cropped video
video_capture.release()
print('Crop video successfully to ==> %s' % self.save_video_path)
"""
保存信息
"""
with open(self.label_path, 'a+') as fil:
"""
save label
"""
fil.write('video-%d, ' % self.video_index)
fil.write(self.save_video_path+', ')
fil.write(self.save_npz_path+', ')
fil.write(str(self.videoFPS))
fil.write(', ')
fil.write(str(self.heartRate))
fil.write('\n')
print('save label ==> %s successfully' % self.label_path)
np.savez(self.save_npz_path, landmarks=landmarksAll, RGB_mean=RGB_mean)
print("save npz ==> %s successfully" % self.save_npz_path)
import sys
import cv2
import numpy as np
sys.path.append('../../GitHub_Face_Toolkit/Peppa_Pig_Face_Engine')
from lib.core.api.facer import FaceAna
from face_alignment.align_trans import get_reference_facial_points, warp_and_crop_face
facer = FaceAna()
landmarks_index = [68, 69, 30, 48, 54] # 68关键点下的5点
crop_version = 'v2'
crop_size = 128
scale = crop_size / 112.0 # 相对于标准尺寸112，放大的倍数
reference = get_reference_facial_points(default_square=True)
if crop_version == 'v2':
reference = (reference - 112.0 / 2) * 1.20 + 112.0 / 2
reference[:, 1] = reference[:, 1] + -20
reference = reference * scale
cap = cv2.VideoCapture(0) # 默认w=640, h=480
# cap.set(cv2.CAP_PROP_FRAME_WIDTH, 960)
# cap.set(cv2.CAP_PROP_FRAME_HEIGHT, 720)
while True:
ret, img_raw = cap.read()
img_h, img_w = img_raw.shape[:2]
import sys
sys.path.append('../Peppa_Pig_Face_Engine-master/')
sys.path.append('../')
from lib.core.api.facer import FaceAna
from lib.core.headpose.pose import get_head_pose, line_pairs
facer = FaceAna()
import torch
import torchvision.transforms as transforms
from face_modules.model import Backbone
from network.AEI_Net import *
import cv2
import PIL.Image as Image
import numpy as np
from face_modules.mtcnn_pytorch.src.align_trans import *
device = torch.device('cuda')
arcface = Backbone(50, 0.6, 'ir_se').to(device)
arcface.eval()
arcface.load_state_dict(torch.load('../face_modules/model_ir_se50.pth', map_location=device), strict=False)
test_transform = transforms.Compose([
transforms.ColorJitter(0.1, 0.1, 0.1, 0.01),
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
])
