def GenerateData(ftxt,data_path,net,argument=False):
'''
:param ftxt: name/path of the text file that contains image path,
bounding box, and landmarks
:param output: path of the output dir
:param net: one of the net in the cascaded networks
:param argument: apply augmentation or not
:return: images and related landmarks
'''
if net == "PNet":
size = 12
elif net == "RNet":
size = 24
elif net == "ONet":
size = 48
else:
print('Net type error')
return
image_id = 0
#
f = open(join(OUTPUT,"landmark_%s_aug.txt" %(size)),'w')
#dstdir = "train_landmark_few"
# get image path , bounding box, and landmarks from file 'ftxt'
data = getDataFromTxt(ftxt,data_path=data_path)
idx = 0
#image_path bbox landmark(5*2)
for (imgPath, bbox, landmarkGt) in data:
#print imgPath
F_imgs = []
F_landmarks = []
print(imgPath)
img = cv2.imread(imgPath)
assert(img is not None)
img_h,img_w,img_c = img.shape
gt_box = np.array([bbox.left,bbox.top,bbox.right,bbox.bottom])
#get sub-image from bbox
f_face = img[bbox.top:bbox.bottom+1,bbox.left:bbox.right+1]
# resize the gt image to specified size
f_face = cv2.resize(f_face,(size,size))
#initialize the landmark
landmark = np.zeros((5, 2))
#normalize land mark by dividing the width and height of the ground truth bounding box
# landmakrGt is a list of tuples
for index, one in enumerate(landmarkGt):
# (( x - bbox.left)/ width of bounding box, (y - bbox.top)/ height of bounding box
rv = ((one[0]-gt_box[0])/(gt_box[2]-gt_box[0]), (one[1]-gt_box[1])/(gt_box[3]-gt_box[1]))
# put the normalized value into the new list landmark
landmark[index] = rv
F_imgs.append(f_face)
F_landmarks.append(landmark.reshape(10))
landmark = np.zeros((5, 2))
if argument:
idx = idx + 1
if idx % 100 == 0:
print(idx, "images done")
x1, y1, x2, y2 = gt_box
#gt's width
gt_w = x2 - x1 + 1
#gt's height
gt_h = y2 - y1 + 1
if max(gt_w, gt_h) < 40 or x1 < 0 or y1 < 0:
continue
#random shift
for i in range(10):
bbox_size = npr.randint(int(min(gt_w, gt_h) * 0.8), np.ceil(1.25 * max(gt_w, gt_h)))
delta_x = npr.randint(-gt_w * 0.2, gt_w * 0.2)
delta_y = npr.randint(-gt_h * 0.2, gt_h * 0.2)
nx1 = int(max(x1+gt_w/2-bbox_size/2+delta_x,0))
ny1 = int(max(y1+gt_h/2-bbox_size/2+delta_y,0))
nx2 = nx1 + bbox_size
ny2 = ny1 + bbox_size
if nx2 > img_w or ny2 > img_h:
continue
crop_box = np.array([nx1,ny1,nx2,ny2])
cropped_im = img[ny1:ny2+1,nx1:nx2+1,:]
resized_im = cv2.resize(cropped_im, (size, size))
#cal iou
iou = IoU(crop_box, np.expand_dims(gt_box,0))
if iou > 0.65:
F_imgs.append(resized_im)
#normalize
for index, one in enumerate(landmarkGt):
rv = ((one[0]-nx1)/bbox_size, (one[1]-ny1)/bbox_size)
landmark[index] = rv
F_landmarks.append(landmark.reshape(10))
landmark = np.zeros((5, 2))
landmark_ = F_landmarks[-1].reshape(-1,2)
bbox = BBox([nx1,ny1,nx2,ny2])
#mirror
if random.choice([0,1]) > 0:
face_flipped, landmark_flipped = flip(resized_im, landmark_)
face_flipped = cv2.resize(face_flipped, (size, size))
#c*h*w
F_imgs.append(face_flipped)
F_landmarks.append(landmark_flipped.reshape(10))
#rotate
if random.choice([0,1]) > 0:
face_rotated_by_alpha, landmark_rotated = rotate(img, bbox, \
bbox.reprojectLandmark(landmark_), 5)#逆时针旋转
#landmark_offset
landmark_rotated = bbox.projectLandmark(landmark_rotated)
face_rotated_by_alpha = cv2.resize(face_rotated_by_alpha, (size, size))
F_imgs.append(face_rotated_by_alpha)
F_landmarks.append(landmark_rotated.reshape(10))
#flip
face_flipped, landmark_flipped = flip(face_rotated_by_alpha, landmark_rotated)
face_flipped = cv2.resize(face_flipped, (size, size))
F_imgs.append(face_flipped)
F_landmarks.append(landmark_flipped.reshape(10))
#anti-clockwise rotation
if random.choice([0,1]) > 0:
face_rotated_by_alpha, landmark_rotated = rotate(img, bbox, \
bbox.reprojectLandmark(landmark_), -5)#顺时针旋转
landmark_rotated = bbox.projectLandmark(landmark_rotated)
face_rotated_by_alpha = cv2.resize(face_rotated_by_alpha, (size, size))
F_imgs.append(face_rotated_by_alpha)
F_landmarks.append(landmark_rotated.reshape(10))
face_flipped, landmark_flipped = flip(face_rotated_by_alpha, landmark_rotated)
face_flipped = cv2.resize(face_flipped, (size, size))
F_imgs.append(face_flipped)
F_landmarks.append(landmark_flipped.reshape(10))
F_imgs, F_landmarks = np.asarray(F_imgs), np.asarray(F_landmarks)
#print F_imgs.shape
#print F_landmarks.shape
for i in range(len(F_imgs)):
#if image_id % 100 == 0:
#print('image id : ', image_id)
if np.sum(np.where(F_landmarks[i] <= 0, 1, 0)) > 0:
continue
if np.sum(np.where(F_landmarks[i] >= 1, 1, 0)) > 0:
continue
cv2.imwrite(join(dstdir,"%d.jpg" %(image_id)), F_imgs[i])
landmarks = map(str,list(F_landmarks[i]))
f.write(join(dstdir,"%d.jpg" %(image_id))+" -2 "+" ".join(landmarks)+"\n")
image_id = image_id + 1
#print F_imgs.shape
#print F_landmarks.shape
#F_imgs = processImage(F_imgs)
#shuffle_in_unison_scary(F_imgs, F_landmarks)
f.close()
return F_imgs,F_landmarks
def GenerateData(ftxt, output,net,argument=False):
if net == "PNet":
size = 12
elif net == "RNet":
size = 24
elif net == "ONet":
size = 48
else:
print('Net type error')
return
image_id = 0
#记录label的txt
f = open(join(OUTPUT,"landmark_%s_aug.txt" %(size)),'w')
data = getDataFromTxt(ftxt)
idx = 0
#image_path bbox landmark(5*2)
for (imgPath, bbox, landmarkGt) in data:
#存储人脸图片和关键点
F_imgs = []
F_landmarks = []
#print(imgPath)
img = cv2.imread(imgPath)
assert(img is not None)
img_h,img_w,img_c = img.shape
gt_box = np.array([bbox.left,bbox.top,bbox.right,bbox.bottom])
#人脸图片
f_face = img[bbox.top:bbox.bottom+1,bbox.left:bbox.right+1]
#resize成网络输入大小
f_face = cv2.resize(f_face,(size,size))
landmark = np.zeros((5, 2))
#normalize
for index, one in enumerate(landmarkGt):
#关键点相对于左上坐标偏移量并归一化
rv = ((one[0]-gt_box[0])/(gt_box[2]-gt_box[0]), (one[1]-gt_box[1])/(gt_box[3]-gt_box[1]))
landmark[index] = rv
F_imgs.append(f_face)
F_landmarks.append(landmark.reshape(10))
landmark = np.zeros((5, 2))
if argument:
#对图像变换
idx = idx + 1
#只取100张照片 节省时间 测试用
if idx % 100 == 0:
print(idx, "images done")
break
x1, y1, x2, y2 = gt_box
#gt's width
gt_w = x2 - x1 + 1
#gt's height
gt_h = y2 - y1 + 1
#除去过小图像
if max(gt_w, gt_h) < 40 or x1 < 0 or y1 < 0:
continue
#random shift,每张照片处理10次
for i in range(10):
#随机裁剪图像大小
bbox_size = npr.randint(int(min(gt_w, gt_h) * 0.8), np.ceil(1.25 * max(gt_w, gt_h)))
#随机左上坐标偏移量
delta_x = npr.randint(-gt_w * 0.2, gt_w * 0.2)
delta_y = npr.randint(-gt_h * 0.2, gt_h * 0.2)
#计算左上坐标
nx1 = max(x1+gt_w/2-bbox_size/2+delta_x,0)
ny1 = max(y1+gt_h/2-bbox_size/2+delta_y,0)
nx2 = nx1 + bbox_size
ny2 = ny1 + bbox_size
#除去超过边界的
if nx2 > img_w or ny2 > img_h:
continue
#裁剪边框，图片
crop_box = np.array([nx1,ny1,nx2,ny2])
nx1, nx2, ny1, ny2 = int(nx1), int(nx2), int(ny1), int(ny2)
cropped_im = img[ny1:ny2+1,nx1:nx2+1,:]
resized_im = cv2.resize(cropped_im, (size, size))
#计算 iou
iou = IoU(crop_box, np.expand_dims(gt_box,0))
#只保留pos图像
if iou > 0.65:
F_imgs.append(resized_im)
#关键点相对偏移 normalize
for index, one in enumerate(landmarkGt):
rv = ((one[0]-nx1)/bbox_size, (one[1]-ny1)/bbox_size)
landmark[index] = rv
F_landmarks.append(landmark.reshape(10))
landmark = np.zeros((5, 2))
landmark_ = F_landmarks[-1].reshape(-1,2)
bbox = BBox([nx1,ny1,nx2,ny2])
#镜像
#mirror
if random.choice([0,1]) > 0:
face_flipped, landmark_flipped = flip(resized_im, landmark_)
face_flipped = cv2.resize(face_flipped, (size, size))
#c*h*w
F_imgs.append(face_flipped)
F_landmarks.append(landmark_flipped.reshape(10))
#逆时针翻转
if random.choice([0,1]) > 0:
face_rotated_by_alpha, landmark_rotated = rotate(img, bbox, \
bbox.reprojectLandmark(landmark_), 5)#逆时针旋转
#关键点偏移
landmark_rotated = bbox.projectLandmark(landmark_rotated)
face_rotated_by_alpha = cv2.resize(face_rotated_by_alpha, (size, size))
F_imgs.append(face_rotated_by_alpha)
F_landmarks.append(landmark_rotated.reshape(10))
#左右翻转
face_flipped, landmark_flipped = flip(face_rotated_by_alpha, landmark_rotated)
face_flipped = cv2.resize(face_flipped, (size, size))
F_imgs.append(face_flipped)
F_landmarks.append(landmark_flipped.reshape(10))
#顺时针翻转
#inverse clockwise rotation
if random.choice([0,1]) > 0:
face_rotated_by_alpha, landmark_rotated = rotate(img, bbox, \
bbox.reprojectLandmark(landmark_), -5)#顺时针旋转
landmark_rotated = bbox.projectLandmark(landmark_rotated)
face_rotated_by_alpha = cv2.resize(face_rotated_by_alpha, (size, size))
F_imgs.append(face_rotated_by_alpha)
F_landmarks.append(landmark_rotated.reshape(10))
face_flipped, landmark_flipped = flip(face_rotated_by_alpha, landmark_rotated)
face_flipped = cv2.resize(face_flipped, (size, size))
F_imgs.append(face_flipped)
F_landmarks.append(landmark_flipped.reshape(10))
F_imgs, F_landmarks = np.asarray(F_imgs), np.asarray(F_landmarks)
##print F_imgs.shape
##print F_landmarks.shape
for i in range(len(F_imgs)):
#print image_id
cv2.imwrite(join(dstdir,"%d.jpg" %(image_id)), F_imgs[i])
landmarks = list(map(str,list(F_landmarks[i])))
f.write(join(dstdir,"%d.jpg" %(image_id))+" -2 "+" ".join(landmarks)+"\n")
image_id = image_id + 1
##print F_imgs.shape
##print F_landmarks.shape
#F_imgs = processImage(F_imgs)
#shuffle_in_unison_scary(F_imgs, F_landmarks)
f.close()
return F_imgs,F_landmarks
def GenerateData(ftxt, output, net, argument=False):
if net == "PNet":
size = 12
elif net == "RNet":
size = 24
elif net == "ONet":
size = 48
else:
print 'Net type error'
return
image_id = 0
f = open(join(OUTPUT, "landmark_%s_aug.txt" % (size)), 'w')
data = getDataFromTxt(ftxt)
idx = 0
#image_path bbox landmark(5*2)
for (imgPath, bbox, landmarkGt) in data:
#print imgPath
F_imgs = []
F_landmarks = []
path = os.path.join(
data_path,
imgPath.split('\\')[0] + '/' + imgPath.split('\\')[1])
img = cv2.imread(path)
assert (img is not None)
img_h, img_w, img_c = img.shape
gt_box = np.array([bbox.left, bbox.top, bbox.right, bbox.bottom])
f_face = img[bbox.top:bbox.bottom + 1, bbox.left:bbox.right + 1]
f_face = cv2.resize(f_face, (size, size))
landmark = np.zeros((5, 2))
#normalize
for index, one in enumerate(landmarkGt):
rv = ((one[0] - gt_box[0]) / (gt_box[2] - gt_box[0]),
(one[1] - gt_box[1]) / (gt_box[3] - gt_box[1]))
landmark[index] = rv
F_imgs.append(f_face)
F_landmarks.append(landmark.reshape(10))
landmark = np.zeros((5, 2))
if argument:
idx = idx + 1
if idx % 100 == 0:
print idx, "images done"
x1, y1, x2, y2 = gt_box
#gt's width
gt_w = x2 - x1 + 1
#gt's height
gt_h = y2 - y1 + 1
if max(gt_w, gt_h) < 40 or x1 < 0 or y1 < 0:
continue
#random shift
for i in range(10):
bbox_size = npr.randint(int(min(gt_w, gt_h) * 0.8),
np.ceil(1.25 * max(gt_w, gt_h)))
delta_x = npr.randint(-gt_w * 0.2, gt_w * 0.2)
delta_y = npr.randint(-gt_h * 0.2, gt_h * 0.2)
nx1 = max(x1 + gt_w / 2 - bbox_size / 2 + delta_x, 0)
ny1 = max(y1 + gt_h / 2 - bbox_size / 2 + delta_y, 0)
nx2 = nx1 + bbox_size
ny2 = ny1 + bbox_size
if nx2 > img_w or ny2 > img_h:
continue
crop_box = np.array([nx1, ny1, nx2, ny2])
cropped_im = img[ny1:ny2 + 1, nx1:nx2 + 1, :]
resized_im = cv2.resize(cropped_im, (size, size))
#cal iou
iou = IoU(crop_box, np.expand_dims(gt_box, 0))
if iou > 0.65:
F_imgs.append(resized_im)
#normalize
for index, one in enumerate(landmarkGt):
rv = ((one[0] - nx1) / bbox_size,
(one[1] - ny1) / bbox_size)
landmark[index] = rv
F_landmarks.append(landmark.reshape(10))
landmark = np.zeros((5, 2))
landmark_ = F_landmarks[-1].reshape(-1, 2)
bbox = BBox([nx1, ny1, nx2, ny2])
# mirror
if random.choice([0, 1]) > 0:
face_flipped, landmark_flipped = flip(
resized_im, landmark_)
face_flipped = cv2.resize(
face_flipped, (size, size),
interpolation=cv2.INTER_LINEAR)
# c*h*w
F_imgs.append(face_flipped)
F_landmarks.append(landmark_flipped.reshape(10))
# rotate
if random.choice([0, 1]) > 0:
face_rotated_by_alpha, landmark_rotated = rotate(
img, bbox, bbox.reprojectLandmark(landmark_),
5) # 逆时针旋转
# landmark_offset
landmark_rotated = bbox.projectLandmark(
landmark_rotated)
face_rotated_by_alpha = cv2.resize(
face_rotated_by_alpha, (size, size),
interpolation=cv2.INTER_LINEAR)
F_imgs.append(face_rotated_by_alpha)
F_landmarks.append(landmark_rotated.reshape(10))
# flip
face_flipped, landmark_flipped = flip(
face_rotated_by_alpha, landmark_rotated)
face_flipped = cv2.resize(
face_flipped, (size, size),
interpolation=cv2.INTER_LINEAR)
F_imgs.append(face_flipped)
F_landmarks.append(landmark_flipped.reshape(10))
# if random.choice([0, 1]) > 0:
# face_rotated_by_alpha, landmark_rotated = rotate(img, bbox,
# bbox.reprojectLandmark(landmark_), 15) # 逆时针旋转
# # landmark_offset
# landmark_rotated = bbox.projectLandmark(landmark_rotated)
# face_rotated_by_alpha = cv2.resize(face_rotated_by_alpha, (size, size), interpolation=cv2.INTER_LINEAR)
# F_imgs.append(face_rotated_by_alpha)
# F_landmarks.append(landmark_rotated.reshape(10))
#
# # flip
# face_flipped, landmark_flipped = flip(face_rotated_by_alpha, landmark_rotated)
# face_flipped = cv2.resize(face_flipped, (size, size), interpolation=cv2.INTER_LINEAR)
# F_imgs.append(face_flipped)
# F_landmarks.append(landmark_flipped.reshape(10))
#
# if random.choice([0, 1]) > 0:
# face_rotated_by_alpha, landmark_rotated = rotate(img, bbox,
# bbox.reprojectLandmark(landmark_), 45) # 逆时针旋转
# # landmark_offset
# landmark_rotated = bbox.projectLandmark(landmark_rotated)
# face_rotated_by_alpha = cv2.resize(face_rotated_by_alpha, (size, size), interpolation=cv2.INTER_LINEAR)
# F_imgs.append(face_rotated_by_alpha)
# F_landmarks.append(landmark_rotated.reshape(10))
#
# # flip
# face_flipped, landmark_flipped = flip(face_rotated_by_alpha, landmark_rotated)
# face_flipped = cv2.resize(face_flipped, (size, size), interpolation=cv2.INTER_LINEAR)
# F_imgs.append(face_flipped)
# F_landmarks.append(landmark_flipped.reshape(10))
#
# if random.choice([0, 1]) > 0:
# face_rotated_by_alpha, landmark_rotated = rotate(img, bbox,
# bbox.reprojectLandmark(landmark_), 90) # 逆时针旋转
# # landmark_offset
# landmark_rotated = bbox.projectLandmark(landmark_rotated)
# face_rotated_by_alpha = cv2.resize(face_rotated_by_alpha, (size, size), interpolation=cv2.INTER_LINEAR)
# F_imgs.append(face_rotated_by_alpha)
# F_landmarks.append(landmark_rotated.reshape(10))
#
# # flip
# face_flipped, landmark_flipped = flip(face_rotated_by_alpha, landmark_rotated)
# face_flipped = cv2.resize(face_flipped, (size, size), interpolation=cv2.INTER_LINEAR)
# F_imgs.append(face_flipped)
# F_landmarks.append(landmark_flipped.reshape(10))
# inverse clockwise rotation
if random.choice([0, 1]) > 0:
face_rotated_by_alpha, landmark_rotated = rotate(
img, bbox, bbox.reprojectLandmark(landmark_),
-5) # 顺时针旋转
landmark_rotated = bbox.projectLandmark(
landmark_rotated)
face_rotated_by_alpha = cv2.resize(
face_rotated_by_alpha, (size, size),
interpolation=cv2.INTER_LINEAR)
F_imgs.append(face_rotated_by_alpha)
F_landmarks.append(landmark_rotated.reshape(10))
face_flipped, landmark_flipped = flip(
face_rotated_by_alpha, landmark_rotated)
face_flipped = cv2.resize(
face_flipped, (size, size),
interpolation=cv2.INTER_LINEAR)
F_imgs.append(face_flipped)
F_landmarks.append(landmark_flipped.reshape(10))
# if random.choice([0, 1]) > 0:
# face_rotated_by_alpha, landmark_rotated = rotate(img, bbox,
# bbox.reprojectLandmark(landmark_),
# -15) # 逆时针旋转
# # landmark_offset
# landmark_rotated = bbox.projectLandmark(landmark_rotated)
# face_rotated_by_alpha = cv2.resize(face_rotated_by_alpha, (size, size), interpolation=cv2.INTER_LINEAR)
# F_imgs.append(face_rotated_by_alpha)
# F_landmarks.append(landmark_rotated.reshape(10))
#
# # flip
# face_flipped, landmark_flipped = flip(face_rotated_by_alpha, landmark_rotated)
# face_flipped = cv2.resize(face_flipped, (size, size), interpolation=cv2.INTER_LINEAR)
# F_imgs.append(face_flipped)
# F_landmarks.append(landmark_flipped.reshape(10))
#
# if random.choice([0, 1]) > 0:
# face_rotated_by_alpha, landmark_rotated = rotate(img, bbox,
# bbox.reprojectLandmark(landmark_),
# -45) # 逆时针旋转
# # landmark_offset
# landmark_rotated = bbox.projectLandmark(landmark_rotated)
# face_rotated_by_alpha = cv2.resize(face_rotated_by_alpha, (size, size), interpolation=cv2.INTER_LINEAR)
# F_imgs.append(face_rotated_by_alpha)
# F_landmarks.append(landmark_rotated.reshape(10))
#
# # flip
# face_flipped, landmark_flipped = flip(face_rotated_by_alpha, landmark_rotated)
# face_flipped = cv2.resize(face_flipped, (size, size), interpolation=cv2.INTER_LINEAR)
# F_imgs.append(face_flipped)
# F_landmarks.append(landmark_flipped.reshape(10))
#
# if random.choice([0, 1]) > 0:
# face_rotated_by_alpha, landmark_rotated = rotate(img, bbox,
# bbox.reprojectLandmark(landmark_),
# -90) # 逆时针旋转
# # landmark_offset
# landmark_rotated = bbox.projectLandmark(landmark_rotated)
# face_rotated_by_alpha = cv2.resize(face_rotated_by_alpha, (size, size), interpolation=cv2.INTER_LINEAR)
# F_imgs.append(face_rotated_by_alpha)
# F_landmarks.append(landmark_rotated.reshape(10))
#
# # flip
# face_flipped, landmark_flipped = flip(face_rotated_by_alpha, landmark_rotated)
# face_flipped = cv2.resize(face_flipped, (size, size), interpolation=cv2.INTER_LINEAR)
# F_imgs.append(face_flipped)
# F_landmarks.append(landmark_flipped.reshape(10))
#
# if random.choice([0, 1]) > 0:
# face_rotated_by_alpha, landmark_rotated = rotate(img, bbox,
# bbox.reprojectLandmark(landmark_),
# 180) # 逆时针旋转
# # landmark_offset
# landmark_rotated = bbox.projectLandmark(landmark_rotated)
# face_rotated_by_alpha = cv2.resize(face_rotated_by_alpha, (size, size), interpolation=cv2.INTER_LINEAR)
# F_imgs.append(face_rotated_by_alpha)
# F_landmarks.append(landmark_rotated.reshape(10))
#
# # flip
# face_flipped, landmark_flipped = flip(face_rotated_by_alpha, landmark_rotated)
# face_flipped = cv2.resize(face_flipped, (size, size), interpolation=cv2.INTER_LINEAR)
# F_imgs.append(face_flipped)
# F_landmarks.append(landmark_flipped.reshape(10))
F_imgs, F_landmarks = np.asarray(F_imgs), np.asarray(F_landmarks)
#print F_imgs.shape
#print F_landmarks.shape
for i in range(len(F_imgs)):
print image_id
if np.sum(np.where(F_landmarks[i] <= 0, 1, 0)) > 0:
continue
if np.sum(np.where(F_landmarks[i] >= 1, 1, 0)) > 0:
continue
cv2.imwrite(join(dstdir, "%d.jpg" % (image_id)), F_imgs[i])
landmarks = map(str, list(F_landmarks[i]))
f.write(
join(dstdir, "%d.jpg" % (image_id)) + " -2 " +
" ".join(landmarks) + "\n")
image_id = image_id + 1
#print F_imgs.shape
#print F_landmarks.shape
#F_imgs = processImage(F_imgs)
#shuffle_in_unison_scary(F_imgs, F_landmarks)
f.close()
return F_imgs, F_landmarks
def GenerateData(ftxt, data_path, net, argument = False):
"""
"""
"""根据网络名来判断size大小"""
if net == 'PNet':
size = 12
elif net == 'RNet':
size = 24
elif net == 'ONet':
size = 48
else:
print('不是PNet，RNet，ONet，net类型错误，请检查')
return
image_id = 0
f = open(os.path.join(OUTPUT, 'landmark_%s_aug.txt'%(size)), 'w')
data = getDataFromTxt(ftxt, data_path = data_path)
#将txt的1234对应BBOx的0123，左上右下
#列表 带有标记位置的路径，label框，标记点坐标
idx = 0#计数图片标记点，label狂的信息提取出了多少
for (imgPath, bbox, landmarkGt) in data:
"""遍历txt文件中的所有数据"""
F_imgs = []#每张照片的所有人脸图像
F_landmarks = []#每张照片的所有标记点
img = cv2.imread(imgPath)
#读取图像路径的图片
assert(img is not None)
img_h, img_w, img_c = img.shape
gt_box = np.array([bbox.left, bbox.top,
bbox.right, bbox.bottom])
#gt_box:左下右上，0123 txt，左右下上
"""作者把上下名称颠倒了"""
f_face = img[bbox.top : bbox.bootom+1, bbox.left : bbox.right+1]
#像素点
f_face = cv2.resize(f_face, (size, size))
#读取人脸图像
landmark = np.zeros((5, 2))
for index, one in enumerate(landmarkGt):
"""遍历每一个图像的5个标记点，
index：0-4
one: 标记点的位置
"""
rv = ((one[0] - gt_box[0]) / (gt_box[2] - gt_box[0]),
(one[1] - gt_box[1]) / (gt_box[3] - gt_box[1]))
#关键点的位置信息转换成比例，距离左下角的距离和边长的比例
landmark[index] = rv
F_imgs.append(f_face)
F_landmarks.append(lanmark.reshap(10))
landmark = np.zeros((5, 2))
if argument:
idx = idx + 1
if idx % 100 == 0:
print(idx, 'images done')
x1, y1, x2, y2 = gt_box
#label框的坐上
gt_w = x2 - x1 + 1
gt_h = y2 - y1 + 1
if max(gt_w, gt_h) < 40 or x1< 0 or y1 < 0:
#如果出错就舍弃这张图片
continue
for i in range(10):
bbox_size = npr.randint(int(min(gt_w, gt_h) * 0.8),
np.ceil(1.25 * max(gt_w, gt_h)))
#偏移量是少20%的最小值，多20%的最大值
delta_x = npr.randint(-gt_w * 0.2, gt_w * 0.2)
delta_y = npr.randint(-gt_h * 0.2, gt_h * 0.2)
nx1 = int(max(x1 + gt_w/2 - bbox_size/2 + delta_x, 0))
ny1 = int(max(y1 + gt_h/2 - bbox_size/2 + delta_y, 0))
#label框随机移动后中心点的位置在，构造label框和移动后
#label框中心重合，但是x1， y1大于0
nx2 = nx1 + bbox_size
ny2 = ny1 + bbox_size
if nx2 > img_w or ny2 > img_h:
#x2,y2还必须在图片内
continue
crop_box = np.array([nx1, ny1, nx2, ny2])
cropped_im = img[ny1 : ny2+1, nx1 : nx2+1,:]
resized_im = cv2.resize(cropped_im, (size, size))
iou = IoU(crop_box, np.expand_dims(gt_box, 0))
if iou > 0.65:
F_imgs.append(resized_im)
#这张图片前期添加了一个label框的像素图像了
#现在添加进去的是构造label框的图像
for index, one in enumerate(landmarkGt):
#将关键点相对于构造label框的信息添加进去
rv = ((one[0] - nx1) / bbox_size,
(one[1] - ny1) / bbox_size)
landmark[index] = rv
F_landmarks.append(landmark.reshape(10))
#这张图片前期添加了一个label框的关键点的信息了
#现在添加进去的是构造label框的关键点的信息了
landmark = np.zeros((5, 2))
landmark_ = Flandmarks[-1].reshape(-1, 2)
bbox = BBox([nx1, ny1, nx2, ny2])
"""随机镜像"""
if random.choice([0, 1]) > 0:
face_flipped, landmark_flipped = \
flip(resized_im, landmark)
face_flipped = cv2.resize(face_flipped,
(size, size))
F_imgs.append(face_flipped)
F_landmarks.append(landmark_flipped.reshape(10))
"""随机反转"""
if random.choice([0, 1]) > 0:
face_roteted_by_alpha, landmark_rotated = \
rotate(img, bbox,
bbox.reprojectLandmark(landmark_),5)
def GenerateData(ftxt, output, net, argument=False):
if net == "PNet":
size = 12
elif net == "RNet":
size = 24
elif net == "ONet":
size = 48
else:
print('Net type error')
return
image_id = 0
f = open(join(OUTPUT, "landmark_%s_aug.txt" % (size)), 'w')
data = getDataFromTxt(
ftxt
) # ftxt == prepare_data/trainImageList.txt， 存储有lfw_5590内保存的landmark信息
idx = 0
for (
imgPath, bbox, landmarkGt
) in data: # 一条data记录：lfw_5590\Aaron_Eckhart_0001.jpg 84 161 92 169 106.25 107.75 146.75 112.25 125.25 142.75 105.25 157.75 139.75 161.75
F_imgs = [] #print imgPath
F_landmarks = []
img = cv2.imread(imgPath)
assert (img is not None)
img_h, img_w, img_c = img.shape
gt_box = np.array([bbox.left, bbox.top, bbox.right, bbox.bottom])
f_face_raw = img[bbox.top:bbox.bottom + 1, bbox.left:bbox.right + 1]
f_face = cv2.resize(f_face_raw, (size, size))
landmark = np.zeros((5, 2))
for index, one in enumerate(landmarkGt): # 归1化处理
rv = ((one[0] - gt_box[0]) / (gt_box[2] - gt_box[0]),
(one[1] - gt_box[1]) / (gt_box[3] - gt_box[1]))
landmark[index] = rv
F_imgs.append(f_face)
F_landmarks.append(landmark.reshape(10))
landmark = np.zeros((5, 2))
if argument:
idx = idx + 1
if idx % 100 == 0:
print(idx, "images done")
x1, y1, x2, y2 = gt_box
gt_w = x2 - x1 + 1 # gt's width
gt_h = y2 - y1 + 1 # gt's height
if max(gt_w, gt_h) < 40 or x1 < 0 or y1 < 0:
continue
for i in range(10): # random shift
bbox_size = npr.randint(int(min(gt_w, gt_h) * 0.8),
np.ceil(1.25 * max(gt_w, gt_h)))
delta_x = npr.randint(-gt_w * 0.2, gt_w * 0.2)
delta_y = npr.randint(-gt_h * 0.2, gt_h * 0.2)
nx1 = max(x1 + gt_w / 2 - bbox_size / 2 + delta_x, 0)
ny1 = max(y1 + gt_h / 2 - bbox_size / 2 + delta_y, 0)
nx2 = nx1 + bbox_size
ny2 = ny1 + bbox_size
if nx2 > img_w or ny2 > img_h:
continue
crop_box = np.array([nx1, ny1, nx2, ny2])
cropped_im = img[int(ny1):int(ny2) + 1,
int(nx1):int(nx2) + 1, :]
resized_im = cv2.resize(cropped_im, (size, size))
iou = IoU(crop_box, np.expand_dims(gt_box, 0)) #cal iou
if iou > 0.65:
F_imgs.append(resized_im)
#normalize
for index, one in enumerate(landmarkGt):
rv = ((one[0] - nx1) / bbox_size,
(one[1] - ny1) / bbox_size)
landmark[index] = rv
F_landmarks.append(landmark.reshape(10))
landmark = np.zeros((5, 2))
landmark_ = F_landmarks[-1].reshape(-1, 2)
bbox = BBox([nx1, ny1, nx2, ny2])
#mirror
if random.choice([0, 1]) > 0:
face_flipped, landmark_flipped = flip(
resized_im, landmark_)
face_flipped = cv2.resize(face_flipped, (size, size))
#c*h*w
F_imgs.append(face_flipped)
F_landmarks.append(landmark_flipped.reshape(10))
#rotate
if random.choice([0, 1]) > 0:
face_rotated_by_alpha, landmark_rotated = rotate(img, bbox, \
bbox.reprojectLandmark(landmark_), 5)#逆时针旋转
#landmark_offset
landmark_rotated = bbox.projectLandmark(
landmark_rotated)
face_rotated_by_alpha = cv2.resize(
face_rotated_by_alpha, (size, size))
F_imgs.append(face_rotated_by_alpha)
F_landmarks.append(landmark_rotated.reshape(10))
#flip
face_flipped, landmark_flipped = flip(
face_rotated_by_alpha, landmark_rotated)
face_flipped = cv2.resize(face_flipped, (size, size))
F_imgs.append(face_flipped)
F_landmarks.append(landmark_flipped.reshape(10))
#inverse clockwise rotation
if random.choice([0, 1]) > 0:
face_rotated_by_alpha, landmark_rotated = rotate(img, bbox, \
bbox.reprojectLandmark(landmark_), -5)#顺时针旋转
landmark_rotated = bbox.projectLandmark(
landmark_rotated)
face_rotated_by_alpha = cv2.resize(
face_rotated_by_alpha, (size, size))
F_imgs.append(face_rotated_by_alpha)
F_landmarks.append(landmark_rotated.reshape(10))
face_flipped, landmark_flipped = flip(
face_rotated_by_alpha, landmark_rotated)
face_flipped = cv2.resize(face_flipped, (size, size))
F_imgs.append(face_flipped)
F_landmarks.append(landmark_flipped.reshape(10))
F_imgs, F_landmarks = np.asarray(F_imgs), np.asarray(F_landmarks)
for i in range(len(F_imgs)):
print(image_id)
if np.sum(np.where(F_landmarks[i] <= 0, 1, 0)) > 0:
continue
if np.sum(np.where(F_landmarks[i] >= 1, 1, 0)) > 0:
continue
cv2.imwrite(join(dstdir, "%d.jpg" % (image_id)), F_imgs[i])
landmarks = map(str, list(F_landmarks[i]))
f.write(
join(dstdir, "%d.jpg" % (image_id)) + " -2 " +
" ".join(landmarks) + "\n")
image_id = image_id + 1
f.close()
return F_imgs, F_landmarks
def main(size):
'''用于处理带有landmark的数据'''
# size=args.input_size
#是否对图像变换
argument = True
if size == 12:
net = 'PNet'
elif size == 24:
net = 'RNet'
elif size == 48:
net = 'ONet'
image_id = 0
#数据输出路径
OUTPUT = os.path.join(data_dir, str(size))
if not os.path.exists(OUTPUT):
os.mkdir(OUTPUT)
#图片处理后输出路径
dstdir = os.path.join(OUTPUT, 'train_%s_landmark_aug' % (net))
if not os.path.exists(dstdir):
os.mkdir(dstdir)
#label记录txt
ftxt = os.path.join(data_dir, 'trainImageList.txt')
#记录label的txt
f = open(os.path.join(OUTPUT, 'landmark_%d_aug.txt' % (size)), 'w')
#获取图像路径，box，关键点
data = getDataFromTxt(ftxt, data_dir)
idx = 0
for (imgPath, box, landmarkGt) in tqdm(data):
#存储人脸图片和关键点
F_imgs = []
F_landmarks = []
img = cv2.imread(imgPath)
img_h, img_w, img_c = img.shape
gt_box = np.array([box.left, box.top, box.right, box.bottom])
#人脸图片
f_face = img[box.top:box.bottom + 1, box.left:box.right + 1]
#resize成网络输入大小
f_face = cv2.resize(f_face, (size, size))
landmark = np.zeros((5, 2))
for index, one in enumerate(landmarkGt):
#关键点相对于左上坐标偏移量并归一化
rv = ((one[0] - gt_box[0]) / (gt_box[2] - gt_box[0]),
(one[1] - gt_box[1]) / (gt_box[3] - gt_box[1]))
landmark[index] = rv
F_imgs.append(f_face)
F_landmarks.append(landmark.reshape(10))
landmark = np.zeros((5, 2))
if argument:
#对图像变换
idx = idx + 1
x1, y1, x2, y2 = gt_box
gt_w = x2 - x1 + 1
gt_h = y2 - y1 + 1
#除去过小图像
if max(gt_w, gt_h) < 40 or x1 < 0 or y1 < 0:
continue
for i in range(10):
#随机裁剪图像大小
box_size = npr.randint(int(min(gt_w, gt_h) * 0.8),
np.ceil(1.25 * max(gt_w, gt_h)))
#随机左上坐标偏移量
delta_x = npr.randint(-gt_w * 0.2, gt_w * 0.2)
delta_y = npr.randint(-gt_h * 0.2, gt_h * 0.2)
#计算左上坐标
nx1 = int(max(x1 + gt_w / 2 - box_size / 2 + delta_x, 0))
ny1 = int(max(y1 + gt_h / 2 - box_size / 2 + delta_y, 0))
nx2 = nx1 + box_size
ny2 = ny1 + box_size
#除去超过边界的
if nx2 > img_w or ny2 > img_h:
continue
#裁剪边框，图片
crop_box = np.array([nx1, ny1, nx2, ny2])
cropped_im = img[ny1:ny2 + 1, nx1:nx2 + 1, :]
resized_im = cv2.resize(cropped_im, (size, size))
iou = IOU(crop_box, np.expand_dims(gt_box, 0))
#只保留pos图像
if iou > 0.65:
F_imgs.append(resized_im)
#关键点相对偏移
for index, one in enumerate(landmarkGt):
rv = ((one[0] - nx1) / box_size,
(one[1] - ny1) / box_size)
landmark[index] = rv
F_landmarks.append(landmark.reshape(10))
landmark = np.zeros((5, 2))
landmark_ = F_landmarks[-1].reshape(-1, 2)
box = BBox([nx1, ny1, nx2, ny2])
#镜像
if random.choice([0, 1]) > 0:
face_flipped, landmark_flipped = flip(
resized_im, landmark_)
face_flipped = cv2.resize(face_flipped, (size, size))
F_imgs.append(face_flipped)
F_landmarks.append(landmark_flipped.reshape(10))
#逆时针翻转
if random.choice([0, 1]) > 0:
face_rotated_by_alpha, landmark_rorated = rotate(
img, box, box.reprojectLandmark(landmark_), 5)
#关键点偏移
landmark_rorated = box.projectLandmark(
landmark_rorated)
face_rotated_by_alpha = cv2.resize(
face_rotated_by_alpha, (size, size))
F_imgs.append(face_rotated_by_alpha)
F_landmarks.append(landmark_rorated.reshape(10))
#左右翻转
face_flipped, landmark_flipped = flip(
face_rotated_by_alpha, landmark_rorated)
face_flipped = cv2.resize(face_flipped, (size, size))
F_imgs.append(face_flipped)
F_landmarks.append(landmark_flipped.reshape(10))
#顺时针翻转
if random.choice([0, 1]) > 0:
face_rotated_by_alpha, landmark_rorated = rotate(
img, box, box.reprojectLandmark(landmark_), -5)
#关键点偏移
landmark_rorated = box.projectLandmark(
landmark_rorated)
face_rotated_by_alpha = cv2.resize(
face_rotated_by_alpha, (size, size))
F_imgs.append(face_rotated_by_alpha)
F_landmarks.append(landmark_rorated.reshape(10))
#左右翻转
face_flipped, landmark_flipped = flip(
face_rotated_by_alpha, landmark_rorated)
face_flipped = cv2.resize(face_flipped, (size, size))
F_imgs.append(face_flipped)
F_landmarks.append(landmark_flipped.reshape(10))
F_imgs, F_landmarks = np.asarray(F_imgs), np.asarray(F_landmarks)
for i in range(len(F_imgs)):
#剔除数据偏移量在[0,1]之间
if np.sum(np.where(F_landmarks[i] <= 0, 1, 0)) > 0:
continue
if np.sum(np.where(F_landmarks[i] >= 1, 1, 0)) > 0:
continue
cv2.imwrite(os.path.join(dstdir, '%d.jpg' % (image_id)), F_imgs[i])
landmarks = list(map(str, list(F_landmarks[i])))
f.write(
os.path.join(dstdir, '%d.jpg' % (image_id)) + ' -2 ' +
' '.join(landmarks) + '\n')
image_id += 1
f.close()
return F_imgs, F_landmarks
def GenerateData(ftxt, output, net, argument=False):
if net == "PNet":
size = 12
elif net == "RNet":
size = 24
elif net == "ONet":
size = 48
else:
print('Net type error')
return
image_id = 0
f = open(
join(OUTPUT, "landmark_%s_aug.txt" % (size)),
'w') # OUTPUT = '12', size = 12 or 24 or 48 根据PNET,RNET,ONET 不同而不同
#dstdir = "train_landmark_few"
# ftxt就是trainImageList.txt 文件
# sample line in the file: lfw_5590/Aaron_Eckhart_0001.jpg 84 161 92 169 106.250000 107.750000 146.750000 112.250000 125.250000 142.750000 105.250000 157.750000 139.750000 161.750000
data = getDataFromTxt(ftxt)
idx = 0
#image_path bbox landmark(5*2)
for (imgPath, bbox, landmarkGt) in data: # 对应上面的sample line 就可以看出数据的结构
#print imgPath
F_imgs = []
F_landmarks = []
img = cv2.imread(imgPath)
if img is None:
continue
img_h, img_w, img_c = img.shape
gt_box = np.array([bbox.left, bbox.top, bbox.right, bbox.bottom])
f_face = img[bbox.top:bbox.bottom + 1, bbox.left:bbox.right + 1]
f_face = cv2.resize(f_face,
(size, size)) # 根据PNET,ONET,RNET 不同这里的size也不同
landmark = np.zeros((5, 2)) # 初始化
#normalize
# （one[0] one[1]）一共5个，每组表示每个landmark的坐标。 (gt_box[0], gt_box[1]) box 左上角坐标， (gt_box[2], gt_box[3]) 表示box 右下脚坐标
for index, one in enumerate(
landmarkGt
): # landmark在train的时候是根据ground truth左上角来的，在没抠出图片之前是根据原图的左上角坐标来的，抠出来以后，又根据抠出来小图的左上角来的
rv = ((one[0] - gt_box[0]) / (gt_box[2] - gt_box[0]),
(one[1] - gt_box[1]) / (gt_box[3] - gt_box[1]))
landmark[index] = rv
F_imgs.append(f_face)
F_landmarks.append(landmark.reshape(10))
landmark = np.zeros((5, 2))
# 对landmark数据增强，因为 WIDER FACE 里人脸检测的trainig data 一共32，032图片并包含393,703人脸，而我们这里landmark traing data 太少 只有10000多张图片
# 其实就是random的平移一下
if argument:
idx = idx + 1
if idx % 100 == 0:
print("%d images done" % idx)
x1, y1, x2, y2 = gt_box
#gt's width
gt_w = x2 - x1 + 1
#gt's height
gt_h = y2 - y1 + 1
if max(gt_w, gt_h) < 40 or x1 < 0 or y1 < 0:
continue
#random shift
for i in range(10):
bbox_size = npr.randint(int(min(gt_w, gt_h) * 0.8),
np.ceil(1.25 * max(gt_w, gt_h)))
delta_x = npr.randint(-gt_w * 0.2, gt_w * 0.2)
delta_y = npr.randint(-gt_h * 0.2, gt_h * 0.2)
nx1 = int(max(x1 + gt_w / 2 - bbox_size / 2 + delta_x, 0))
ny1 = int(max(y1 + gt_h / 2 - bbox_size / 2 + delta_y, 0))
nx2 = nx1 + bbox_size
ny2 = ny1 + bbox_size
if nx2 > img_w or ny2 > img_h:
continue
crop_box = np.array([nx1, ny1, nx2, ny2])
cropped_im = img[ny1:ny2 + 1, nx1:nx2 + 1, :]
resized_im = cv2.resize(cropped_im, (size, size))
#cal iou
iou = IoU(crop_box, np.expand_dims(gt_box, 0))
if iou > 0.65: # 如果平移完了，与ground truth iou还是大于 0.65，我们认为还是landmark的样本
F_imgs.append(resized_im)
#normalize
for index, one in enumerate(landmarkGt):
rv = ((one[0] - nx1) / bbox_size,
(one[1] - ny1) / bbox_size)
landmark[index] = rv
F_landmarks.append(landmark.reshape(10))
landmark = np.zeros((5, 2))
landmark_ = F_landmarks[-1].reshape(-1, 2)
bbox = BBox([nx1, ny1, nx2, ny2])
#mirror i.e. 左右翻转一下图片
if random.choice([0, 1]) > 0:
face_flipped, landmark_flipped = flip(
resized_im, landmark_)
face_flipped = cv2.resize(face_flipped, (size, size))
#c*h*w
F_imgs.append(face_flipped)
F_landmarks.append(landmark_flipped.reshape(10))
#rotate i.e. 随机旋转
if random.choice([0, 1]) > 0:
face_rotated_by_alpha, landmark_rotated = rotate(img, bbox, \
bbox.reprojectLandmark(landmark_), 5)#逆时针旋转
#landmark_offset
landmark_rotated = bbox.projectLandmark(
landmark_rotated)
face_rotated_by_alpha = cv2.resize(
face_rotated_by_alpha, (size, size))
F_imgs.append(face_rotated_by_alpha)
F_landmarks.append(landmark_rotated.reshape(10))
#flip
face_flipped, landmark_flipped = flip(
face_rotated_by_alpha, landmark_rotated)
face_flipped = cv2.resize(face_flipped, (size, size))
F_imgs.append(face_flipped)
F_landmarks.append(landmark_flipped.reshape(10))
#inverse clockwise rotation
if random.choice([0, 1]) > 0:
face_rotated_by_alpha, landmark_rotated = rotate(img, bbox, \
bbox.reprojectLandmark(landmark_), -5)#顺时针旋转
landmark_rotated = bbox.projectLandmark(
landmark_rotated)
face_rotated_by_alpha = cv2.resize(
face_rotated_by_alpha, (size, size))
F_imgs.append(face_rotated_by_alpha)
F_landmarks.append(landmark_rotated.reshape(10))
face_flipped, landmark_flipped = flip(
face_rotated_by_alpha, landmark_rotated)
face_flipped = cv2.resize(face_flipped, (size, size))
F_imgs.append(face_flipped)
F_landmarks.append(landmark_flipped.reshape(10))
F_imgs, F_landmarks = np.asarray(F_imgs), np.asarray(F_landmarks)
#print F_imgs.shape
#print F_landmarks.shape
for i in range(len(F_imgs)):
if np.sum(np.where(F_landmarks[i] <= 0, 1, 0)) > 0:
continue
if np.sum(np.where(F_landmarks[i] >= 1, 1, 0)) > 0:
continue
cv2.imwrite(join(dstdir, "%d.jpg" % (image_id)), F_imgs[i])
landmarks = map(str, list(F_landmarks[i]))
f.write(
join(dstdir, "%d.jpg" % (image_id)) + " -2 " +
" ".join(landmarks) + "\n")
image_id = image_id + 1
#print F_imgs.shape
#print F_landmarks.shape
#F_imgs = processImage(F_imgs)
#shuffle_in_unison_scary(F_imgs, F_landmarks)
f.close()
return F_imgs, F_landmarks
def GenerateData(ftxt, output,net,argument=False):
if net == "PNet":
size = 12
elif net == "RNet":
size = 24
elif net == "ONet":
size = 48
else:
print 'Net type error'
return
image_id = 0
f = open(join(OUTPUT,"landmark_%s_aug.txt" %(size)),'w')
#dstdir = "train_landmark_few"
data = getDataFromTxt(ftxt)
idx = 0
#image_path bbox landmark(5*2)
for (imgPath, bbox, landmarkGt) in data:
#print imgPath
F_imgs = []
F_landmarks = []
img = cv2.imread(imgPath)
assert(img is not None)
img_h,img_w,img_c = img.shape
gt_box = np.array([bbox.left,bbox.top,bbox.right,bbox.bottom])
f_face = img[bbox.top:bbox.bottom+1,bbox.left:bbox.right+1]
f_face = cv2.resize(f_face,(size,size))
landmark = np.zeros((5, 2))
#normalize
for index, one in enumerate(landmarkGt):
rv = ((one[0]-gt_box[0])/(gt_box[2]-gt_box[0]), (one[1]-gt_box[1])/(gt_box[3]-gt_box[1]))
landmark[index] = rv
F_imgs.append(f_face)
F_landmarks.append(landmark.reshape(10))
landmark = np.zeros((5, 2))
if argument:
idx = idx + 1
if idx % 100 == 0:
print idx, "images done"
x1, y1, x2, y2 = gt_box
#gt's width
gt_w = x2 - x1 + 1
#gt's height
gt_h = y2 - y1 + 1
if max(gt_w, gt_h) < 40 or x1 < 0 or y1 < 0:
continue
#random shift
for i in range(10):
bbox_size = npr.randint(int(min(gt_w, gt_h) * 0.8), np.ceil(1.25 * max(gt_w, gt_h)))
delta_x = npr.randint(-gt_w * 0.2, gt_w * 0.2)
delta_y = npr.randint(-gt_h * 0.2, gt_h * 0.2)
nx1 = max(x1+gt_w/2-bbox_size/2+delta_x,0)
ny1 = max(y1+gt_h/2-bbox_size/2+delta_y,0)
nx2 = nx1 + bbox_size
ny2 = ny1 + bbox_size
if nx2 > img_w or ny2 > img_h:
continue
crop_box = np.array([nx1,ny1,nx2,ny2])
cropped_im = img[ny1:ny2+1,nx1:nx2+1,:]
resized_im = cv2.resize(cropped_im, (size, size))
#cal iou
iou = IoU(crop_box, np.expand_dims(gt_box,0))
if iou > 0.65:
F_imgs.append(resized_im)
#normalize
for index, one in enumerate(landmarkGt):
rv = ((one[0]-nx1)/bbox_size, (one[1]-ny1)/bbox_size)
landmark[index] = rv
F_landmarks.append(landmark.reshape(10))
landmark = np.zeros((5, 2))
landmark_ = F_landmarks[-1].reshape(-1,2)
bbox = BBox([nx1,ny1,nx2,ny2])
#mirror
if random.choice([0,1]) > 0:
face_flipped, landmark_flipped = flip(resized_im, landmark_)
face_flipped = cv2.resize(face_flipped, (size, size))
#c*h*w
F_imgs.append(face_flipped)
F_landmarks.append(landmark_flipped.reshape(10))
#rotate
if random.choice([0,1]) > 0:
face_rotated_by_alpha, landmark_rotated = rotate(img, bbox, \
bbox.reprojectLandmark(landmark_), 5)#逆时针旋转
#landmark_offset
landmark_rotated = bbox.projectLandmark(landmark_rotated)
face_rotated_by_alpha = cv2.resize(face_rotated_by_alpha, (size, size))
F_imgs.append(face_rotated_by_alpha)
F_landmarks.append(landmark_rotated.reshape(10))
#flip
face_flipped, landmark_flipped = flip(face_rotated_by_alpha, landmark_rotated)
face_flipped = cv2.resize(face_flipped, (size, size))
F_imgs.append(face_flipped)
F_landmarks.append(landmark_flipped.reshape(10))
#inverse clockwise rotation
if random.choice([0,1]) > 0:
face_rotated_by_alpha, landmark_rotated = rotate(img, bbox, \
bbox.reprojectLandmark(landmark_), -5)#顺时针旋转
landmark_rotated = bbox.projectLandmark(landmark_rotated)
face_rotated_by_alpha = cv2.resize(face_rotated_by_alpha, (size, size))
F_imgs.append(face_rotated_by_alpha)
F_landmarks.append(landmark_rotated.reshape(10))
face_flipped, landmark_flipped = flip(face_rotated_by_alpha, landmark_rotated)
face_flipped = cv2.resize(face_flipped, (size, size))
F_imgs.append(face_flipped)
F_landmarks.append(landmark_flipped.reshape(10))
F_imgs, F_landmarks = np.asarray(F_imgs), np.asarray(F_landmarks)
#print F_imgs.shape
#print F_landmarks.shape
for i in range(len(F_imgs)):
print image_id
if np.sum(np.where(F_landmarks[i] <= 0, 1, 0)) > 0:
continue
if np.sum(np.where(F_landmarks[i] >= 1, 1, 0)) > 0:
continue
cv2.imwrite(join(dstdir,"%d.jpg" %(image_id)), F_imgs[i])
landmarks = map(str,list(F_landmarks[i]))
f.write(join(dstdir,"%d.jpg" %(image_id))+" -2 "+" ".join(landmarks)+"\n")
image_id = image_id + 1
#print F_imgs.shape
#print F_landmarks.shape
#F_imgs = processImage(F_imgs)
#shuffle_in_unison_scary(F_imgs, F_landmarks)
f.close()
return F_imgs,F_landmarks
def GenerateData(ftxt, output, net, argument=False):
if net == 'PNet':
size = 12
elif net == 'RNet':
size = 24
elif net == 'ONet':
size = 48
else:
return
image_id = 0
f = open(join(OUTPUT, "landmark_%s_aug.txt" % (size)), 'w')
data = getDataFromTxt(ftxt) #获取注解文件信息
idx = 0
for (imgPath, bbox, landmarkGt) in data:
F_imgs = [] #存储图片信息
F_landmarks = [] #存储坐标
img = cv2.imread(imgPath)
img_h, img_w, img_c = img.shape
gt_box = np.array([bbox.left, bbox.top, bbox.right,
bbox.bottom]) #获取真实的包围框
f_face = img[bbox.top:bbox.bottom + 1, bbox.left:bbox.right + 1]
f_face = cv2.resize(f_face, (size, size)) #截取人脸图片并转换尺寸
landmark = np.zeros((5, 2))
#标准化
for index, one in enumerate(landmarkGt):
rv = ((one[0] - gt_box[0]) / (gt_box[2] - gt_box[0]),
(one[1] - gt_box[1]) / (gt_box[3] - gt_box[1])) #规划到0-1之间
landmark[index] = rv
F_imgs.append(f_face)
F_landmarks.append(landmark.reshape(10))
if argument:
idx = idx + 1
x1, y1, x2, y2 = gt_box
gt_w = x2 - x1 + 1
gt_h = y2 - y1 + 1
if max(gt_w, gt_h) < 40 or x1 < 0 or y1 < 0:
continue
# random shift
for i in range(10):
bbox_size = npr.randint(int(min(gt_w, gt_h) * 0.8),
np.ceil(1.25 * max(gt_w, gt_h)))
delta_x = npr.randint(-gt_w * 0.2, gt_w * 0.2)
delta_y = npr.randint(-gt_h * 0.2, gt_h * 0.2)
nx1 = int(max(x1 + gt_w / 2 - bbox_size / 2 + delta_x, 0))
ny1 = int(max(y1 + gt_h / 2 - bbox_size / 2 + delta_y, 0))
nx2 = nx1 + bbox_size
ny2 = ny1 + bbox_size
if nx2 > img_w or ny2 > img_h:
continue
crop_box = np.array([nx1, ny1, nx2, ny2])
cropped_im = img[ny1:ny2 + 1, nx1:nx2 + 1, :]
resized_im = cv2.resize(cropped_im, (size, size))
iou = IoU(crop_box, np.expand_dims(gt_box, 0))
if iou > 0.65:
F_imgs.append(resized_im)
for index, one in enumerate(landmarkGt):
rv = ((one[0] - nx1) / bbox_size,
(one[1] - ny1) / bbox_size)
landmark[index] = rv
F_landmarks.append(landmark.reshape(10))
landmark = np.zeros((5, 2))
landmark_ = F_landmarks[-1].reshape(-1, 2)
bbox = BBox([nx1, ny1, nx2, ny2])
if random.choice([0, 1]) > 0:
face_flipped, landmark_flipped = flip(
resized_im, landmark_)
face_flipped = cv2.resize(face_flipped, (size, size))
F_imgs.append(face_flipped)
F_landmarks.append(landmark_flipped.reshape(10))
if random.choice([0, 1]) > 0:
face_rotated_by_alpha, landmark_rotated = rotate(img, bbox, \
bbox.reprojectLandmark(landmark_), 5) # 逆时针旋转
landmark_rotated = bbox.projectLandmark(
landmark_rotated)
face_rotated_by_alpha = cv2.resize(
face_rotated_by_alpha, (size, size))
F_imgs.append(face_rotated_by_alpha)
F_landmarks.append(landmark_rotated.reshape(10))
face_flipped, landmark_flipped = flip(
face_rotated_by_alpha, landmark_rotated)
face_flipped = cv2.resize(face_flipped, (size, size))
F_imgs.append(face_flipped)
F_landmarks.append(landmark_flipped.reshape(10))
if random.choice([0, 1]) > 0:
face_rotated_by_alpha, landmark_rotated = rotate(img, bbox, \
bbox.reprojectLandmark(landmark_), -5) # 顺时针旋转
landmark_rotated = bbox.projectLandmark(
landmark_rotated)
face_rotated_by_alpha = cv2.resize(
face_rotated_by_alpha, (size, size))
F_imgs.append(face_rotated_by_alpha)
F_landmarks.append(landmark_rotated.reshape(10))
face_flipped, landmark_flipped = flip(
face_rotated_by_alpha, landmark_rotated)
face_flipped = cv2.resize(face_flipped, (size, size))
F_imgs.append(face_flipped)
F_landmarks.append(landmark_flipped.reshape(10))
F_imgs, F_landmarks = np.asarray(F_imgs), np.asarray(F_landmarks)
for i in range(len(F_imgs)):
if np.sum(np.where(F_landmarks[i] <= 0, 1, 0)) > 0:
continue
if np.sum(np.where(F_landmarks[i] >= 1, 1, 0)) > 0:
continue
cv2.imwrite(join(dstdir, "%d.jpg" % (image_id)), F_imgs[i])
landmarks = map(str, list(F_landmarks[i]))
f.write(
join(dstdir, "%d.jpg" % (image_id)) + " -2 " +
" ".join(landmarks) + "\n")
image_id = image_id + 1
f.close()
return F_imgs, F_landmarks
def GenerateData(ftxt, output,net,argument=False):
if net == "PNet":
size = 12
elif net == "RNet":
size = 24
elif net == "ONet":
size = 48
else:
print('Net type error')
return
image_id = 0
f = open(join(OUTPUT,"landmark_%s_aug.txt" %(size)), 'w')
#dstdir = "train_landmark_few"
data = getDataFromTxt(ftxt) # ftxt == trainImageList.txt，从文本里面拿到每张图片人脸的3份数据（文件名，人脸框，特征点坐标 ）
idx = 0
for (imgPath, bbox, landmarkGt) in data: #遍历每张原始照片，image_path bbox landmark(5*2)
F_imgs = []
F_landmarks = []
img = cv2.imread(imgPath)
assert(img is not None)
img_h, img_w, img_c = img.shape
gt_box = np.array([bbox.left, bbox.top, bbox.right, bbox.bottom]) # 标签中人脸的区域
f_face = img[bbox.top: bbox.bottom+1, bbox.left: bbox.right+1] # 拿到标签中人脸的矩形框
f_face = cv2.resize(f_face, (size, size)) # 将人脸resize成12 * 12大小
landmark = np.zeros((5, 2))
#normalize
for index, one in enumerate(landmarkGt): # 遍历人脸5个特征点的label,index是数组下标。
rv = ((one[0]-gt_box[0])/(gt_box[2]-gt_box[0]), (one[1]-gt_box[1])/(gt_box[3]-gt_box[1])) # rv是一个（x, y），标识1个特征点坐标，做归一化处理，相对于label框框的比例
landmark[index] = rv
F_imgs.append(f_face) # 将1个人脸像素数据加入F_img的list中
F_landmarks.append(landmark.reshape(10)) # 将1个人脸特征点数据加入F_landmarks的list中
landmark = np.zeros((5, 2))
if argument:
idx = idx + 1
if idx % 100 == 0:
print(idx, "images done")
x1, y1, x2, y2 = gt_box # 点1（x1，y1）位于左上角，点2（x2，y2）位于右下角
gt_w = x2 - x1 + 1 # gt's width
gt_h = y2 - y1 + 1 # gt's height
if max(gt_w, gt_h) < 40 or x1 < 0 or y1 < 0: # 过小的图片 && 非法图片忽略
continue
#random shift
for i in range(10):
bbox_size = npr.randint(int(min(gt_w, gt_h) * 0.8), np.ceil(1.25 * max(gt_w, gt_h)))
delta_x = npr.randint(-gt_w * 0.2, gt_w * 0.2)
delta_y = npr.randint(-gt_h * 0.2, gt_h * 0.2)
nx1 = max(x1+gt_w/2-bbox_size/2+delta_x,0) # gt_w/2-bbox_size/2 == 因为box尺寸变化产生的偏移
ny1 = max(y1+gt_h/2-bbox_size/2+delta_y,0)
nx2 = nx1 + bbox_size
ny2 = ny1 + bbox_size
if nx2 > img_w or ny2 > img_h:
continue
crop_box = np.array([nx1, ny1, nx2, ny2])
cropped_im = img[int(ny1):int(ny2)+1, int(nx1):int(nx2)+1, :]
resized_im = cv2.resize(cropped_im, (size, size))
#cal iou
iou = IoU(crop_box, np.expand_dims(gt_box, 0))
if iou > 0.65:
F_imgs.append(resized_im) # 将12*12的crop框框也加入F_imgs 的list中
for index, one in enumerate(landmarkGt): # normalize归一化处理
rv = ((one[0]-nx1)/bbox_size, (one[1]-ny1)/bbox_size)
landmark[index] = rv
F_landmarks.append(landmark.reshape(10))
landmark = np.zeros((5, 2))
landmark_ = F_landmarks[-1].reshape(-1,2) # 将最新加入的list[-1]的len为10的一维数组，reshape成(-1,2)
bbox = BBox([nx1, ny1, nx2, ny2])
# 1 mirror # 翻转，以轴为参照
if random.choice([0,1]) > 0: #
face_flipped, landmark_flipped = flip(resized_im, landmark_) # 截取的像素数据 && 特征点数据做翻转
face_flipped = cv2.resize(face_flipped, (size, size))
#c*h*w
F_imgs.append(face_flipped) # 12 * 12的图片加入F_imgs队列。
F_landmarks.append(landmark_flipped.reshape(10))
# 逆时针旋转 以中心点为参照, 旋转α角度
if random.choice([0, 1]) > 0:
face_rotated_by_alpha, landmark_rotated = rotate(img, bbox, \
bbox.reprojectLandmark(landmark_), 5)#逆时针旋转
# 2 顺时针旋转
landmark_rotated = bbox.projectLandmark(landmark_rotated) # 根据landmark的坐标做重新生成landmark_rotated？
face_rotated_by_alpha = cv2.resize(face_rotated_by_alpha, (size, size))
F_imgs.append(face_rotated_by_alpha)
F_landmarks.append(landmark_rotated.reshape(10))
# 3 顺时针旋转后再翻转
face_flipped, landmark_flipped = flip(face_rotated_by_alpha, landmark_rotated)
face_flipped = cv2.resize(face_flipped, (size, size))
F_imgs.append(face_flipped)
F_landmarks.append(landmark_flipped.reshape(10))
if random.choice([0,1]) > 0:
# 4 逆时针旋转
face_rotated_by_alpha, landmark_rotated = rotate(img, bbox, \
bbox.reprojectLandmark(landmark_), -5)#顺时针旋转
landmark_rotated = bbox.projectLandmark(landmark_rotated)
face_rotated_by_alpha = cv2.resize(face_rotated_by_alpha, (size, size))
F_imgs.append(face_rotated_by_alpha)
F_landmarks.append(landmark_rotated.reshape(10))
# 5 逆时针旋转
face_flipped, landmark_flipped = flip(face_rotated_by_alpha, landmark_rotated)
face_flipped = cv2.resize(face_flipped, (size, size))
F_imgs.append(face_flipped)
F_landmarks.append(landmark_flipped.reshape(10))
F_imgs, F_landmarks = np.asarray(F_imgs), np.asarray(F_landmarks) # 将list转换为数组，维度保持不变。
#print F_imgs.shape # print F_landmarks.shape
for i in range(len(F_imgs)): # 一张pic里面，数据增强后形成所有 12 * 12的图片
print(image_id) # 剔除非法特征点？哪里来的非法特征点？
if np.sum(np.where(F_landmarks[i] <= 0, 1, 0)) > 0: # 将F_landmarks中所有 <= 0的数字变为1，>0 的数变为0
continue
if np.sum(np.where(F_landmarks[i] >= 1, 1, 0)) > 0:
continue
cv2.imwrite(join(dstdir,"%d.jpg" %(image_id)), F_imgs[i]) # dstdir=12/train_PNet_landmark_aug"
landmarks = map(str, list(F_landmarks[i]))
f.write(join(dstdir, "%d.jpg" %(image_id))+" -2 "+" ".join(landmarks)+"\n") # 将特征点坐标，写入landmark_12_aug的txt文件。 12/train_PNet_landmark_aug
image_id = image_id + 1
f.close()
return F_imgs,F_landmarks
def GenerateData(ftxt, fdir, output, net, dstdir, argument=False):
if net == "PNet":
size = 12
elif net == "RNet":
size = 24
elif net == "ONet":
size = 48
else:
print 'Net type error'
return
image_id = 0
f = open(join(output, "landmark.txt"), 'w')
data = getDataFromTxt(ftxt, fdir)
idx = 0
#image_path bbox landmark(5*2)
for (imgPath, bbox, landmarkGt) in data:
#print imgPath
F_imgs = []
F_bbox = []
F_landmarks = []
img = cv2.imread(imgPath)
print imgPath
assert (img is not None)
img_h, img_w, img_c = img.shape
gt_box = np.array([bbox.left, bbox.top, bbox.right, bbox.bottom])
'''
##########need commented
f_face = img[bbox.top:bbox.bottom+1,bbox.left:bbox.right+1]
f_face = cv2.resize(f_face,(size,size))
landmark = np.zeros((5, 2))
#normalize
for index, one in enumerate(landmarkGt):
rv = ((one[0]-gt_box[0])/(gt_box[2]-gt_box[0]), (one[1]-gt_box[1])/(gt_box[3]-gt_box[1]))
landmark[index] = rv
F_imgs.append(f_face)
F_landmarks.append(landmark.reshape(10))
#show_landmark(f_face,landmark)
############
'''
landmark = np.zeros((5, 2))
if argument:
idx = idx + 1
if idx % 100 == 0:
print idx, "images done"
x1, y1, x2, y2 = gt_box
#gt's width
gt_w = x2 - x1 + 1
#gt's height
gt_h = y2 - y1 + 1
if max(gt_w, gt_h) < 40 or x1 < 0 or y1 < 0:
continue
#random shift
for i in range(15):
bbox_size = npr.randint(int(min(gt_w, gt_h) * 0.8),
np.ceil(1.25 * max(gt_w, gt_h)))
delta_x = npr.randint(-gt_w * 0.2, gt_w * 0.2)
delta_y = npr.randint(-gt_h * 0.2, gt_h * 0.2)
nx1 = max(x1 + gt_w / 2 - bbox_size / 2 + delta_x, 0)
ny1 = max(y1 + gt_h / 2 - bbox_size / 2 + delta_y, 0)
nx2 = nx1 + bbox_size
ny2 = ny1 + bbox_size
if nx2 > img_w or ny2 > img_h:
continue
crop_box = np.array([nx1, ny1, nx2, ny2])
cropped_im = img[ny1:ny2 + 1, nx1:nx2 + 1, :]
resized_im = cv2.resize(cropped_im, (size, size))
#cal iou
iou = IoU(crop_box, np.expand_dims(gt_box, 0))
if iou > 0.65:
F_imgs.append(resized_im)
offset_x1 = (x1 - nx1) / float(gt_w)
offset_y1 = (y1 - ny1) / float(gt_h)
offset_x2 = (x2 - nx2) / float(gt_w)
offset_y2 = (y2 - ny2) / float(gt_h)
#normalize
for index, one in enumerate(landmarkGt):
rv = ((one[0] - nx1) / bbox_size,
(one[1] - ny1) / bbox_size)
landmark[index] = rv
#F_bbox.append([offset_x1, offset_y1, offset_x2, offset_y2])
F_landmarks.append(landmark.reshape(10))
#print landmark
#print 'landmark\n',landmark.reshape(10)
landmark_ = landmark.copy()
#show_landmark(resized_im,landmark)
landmark = np.zeros((5, 2))
#landmark_ = F_landmarks[-1].reshape(-1,2)
bbox = BBox([nx1, ny1, nx2, ny2])
#mirror
if random.choice([0,1]) > 0 or (landmark_[0][0]+landmark_[3][0])>=2*landmark_[2][0] \
or (landmark_[1][0]+landmark_[4][0])>=2*landmark_[2][0]:
face_flipped, landmark_flipped = flip(
resized_im, landmark_)
face_flipped = cv2.resize(face_flipped, (size, size))
#c*h*w
F_imgs.append(face_flipped)
#F_bbox.append([-offset_x2, offset_y1, -offset_x1, offset_y2])
F_landmarks.append(landmark_flipped.reshape(10))
#show_landmark(face_flipped,landmark_flipped)
#rotate
if random.choice([0,1]) > 0 or (landmark_[0][0]+landmark_[3][0])>=2*landmark_[2][0] \
or (landmark_[1][0]+landmark_[4][0])>=2*landmark_[2][0]:
face_rotated_by_alpha, landmark_rotated = rotate(img, bbox, \
landmarkGt, 20)#逆时针旋转
#landmark_offset
landmark_rotated = bbox.projectLandmark(
landmark_rotated)
face_rotated_by_alpha = cv2.resize(
face_rotated_by_alpha, (size, size))
F_imgs.append(face_rotated_by_alpha)
F_landmarks.append(landmark_rotated.reshape(10))
#show_landmark(face_rotated_by_alpha,landmark_rotated)
#flip
face_flipped, landmark_flipped = flip(
face_rotated_by_alpha, landmark_rotated)
face_flipped = cv2.resize(face_flipped, (size, size))
F_imgs.append(face_flipped)
F_landmarks.append(landmark_flipped.reshape(10))
#show_landmark(face_flipped,landmark_flipped)
#inverse clockwise rotation
if random.choice([0,1]) > 0 or (landmark_[0][0]+landmark_[3][0])>=2*landmark_[2][0] \
or (landmark_[1][0]+landmark_[4][0])>=2*landmark_[2][0]:
face_rotated_by_alpha, landmark_rotated = rotate(img, bbox, \
landmarkGt, -20)#顺时针旋转
landmark_rotated = bbox.projectLandmark(
landmark_rotated)
face_rotated_by_alpha = cv2.resize(
face_rotated_by_alpha, (size, size))
F_imgs.append(face_rotated_by_alpha)
F_landmarks.append(landmark_rotated.reshape(10))
#show_landmark(face_rotated_by_alpha,landmark_rotated)
face_flipped, landmark_flipped = flip(
face_rotated_by_alpha, landmark_rotated)
face_flipped = cv2.resize(face_flipped, (size, size))
F_imgs.append(face_flipped)
F_landmarks.append(landmark_flipped.reshape(10))
#show_landmark(face_flipped,landmark_flipped)
F_imgs, F_landmarks = np.asarray(F_imgs), np.asarray(F_landmarks)
#print F_imgs.shape
#print F_landmarks.shape
for i in range(len(F_imgs)):
if np.sum(np.where(F_landmarks[i] <= 0, 1, 0)) > 0:
continue
if np.sum(np.where(F_landmarks[i] >= 1, 1, 0)) > 0:
continue
cv2.imwrite(join(dstdir, "%d.jpg" % (image_id)), F_imgs[i])
str_landmark = F_landmarks[i].copy()
str_landmark[0] = F_landmarks[i, 0]
str_landmark[1] = F_landmarks[i, 2]
str_landmark[2] = F_landmarks[i, 4]
str_landmark[3] = F_landmarks[i, 6]
str_landmark[4] = F_landmarks[i, 8]
str_landmark[5] = F_landmarks[i, 1]
str_landmark[6] = F_landmarks[i, 3]
str_landmark[7] = F_landmarks[i, 5]
str_landmark[8] = F_landmarks[i, 7]
str_landmark[9] = F_landmarks[i, 9]
landmarks = map(str, list(str_landmark))
f.write(
join(dstdir, "%d.jpg" % (image_id)) + " -1 -1 -1 -1 -1 " +
" ".join(landmarks) + "\n")
image_id = image_id + 1
f.close()
return F_imgs, F_landmarks
net = 'ONet'
image_id = 0
#数据输出路径
OUTPUT = os.path.join(data_dir, str(size))
if not os.path.exists(OUTPUT):
os.makedirs(OUTPUT)
#图片处理后输出路径
dstdir = os.path.join(OUTPUT, 'train_%s_landmark_aug' % (net))
if not os.path.exists(dstdir):
os.makedirs(dstdir)
#label记录txt
ftxt = os.path.join(data_dir, 'trainImageList_all_celeba.txt')
#记录label的txt
f = open(os.path.join(OUTPUT, 'landmark_%d_aug.txt' % (size)), 'w')
#获取图像路径，box，关键点
data = getDataFromTxt(ftxt, data_dir)
idx = 0
for (imgPath, box, landmarkGt) in tqdm(data):
#存储人脸图片和关键点
F_imgs = []
F_landmarks = []
img = cv2.imread(imgPath)
img_h, img_w, img_c = img.shape
gt_box = np.array([box.left, box.top, box.right, box.bottom])
#人脸图片
f_face = img[box.top:box.bottom + 1, box.left:box.right + 1]
if not np.array(f_face.shape).all():
continue