def __getitem__(self, index):
img = Image.open(os.path.join(self.data_dir, self.X_train[index] + self.img_ext))
img = img.convert(self.image_mode)
mat_path = os.path.join(self.data_dir, self.y_train[index] + self.annot_ext)
# Crop the face loosely
pt2d = utils.get_pt2d_from_mat(mat_path)
x_min = min(pt2d[0, :])
y_min = min(pt2d[1, :])
x_max = max(pt2d[0, :])
y_max = max(pt2d[1, :])
k = 0.20
x_min -= 2 * k * abs(x_max - x_min)
y_min -= 2 * k * abs(y_max - y_min)
x_max += 2 * k * abs(x_max - x_min)
y_max += 0.6 * k * abs(y_max - y_min)
img = img.crop((int(x_min), int(y_min), int(x_max), int(y_max)))
# We get the pose in radians
pose = utils.get_ypr_from_mat(mat_path)
# And convert to degrees.
pitch = pose[0] * 180 / np.pi
yaw = pose[1] * 180 / np.pi
roll = pose[2] * 180 / np.pi
# Bin values
bins = np.array(range(-99, 102, 3))
labels = torch.LongTensor(np.digitize([yaw, pitch, roll], bins) - 1)
cont_labels = torch.FloatTensor([yaw, pitch, roll])
if self.transform is not None:
img = self.transform(img)
return img, labels, cont_labels, self.X_train[index]
def __getitem__(self, index):
img = Image.open(
os.path.join(self.data_dir, self.X_train[index] + self.img_ext))
img = img.convert(self.image_mode)
mat_path = os.path.join(self.data_dir,
self.y_train[index] + self.annot_ext)
# Crop the face loosely
pt2d = utils.get_pt2d_from_mat(mat_path)
x_min = min(pt2d[0, :])
y_min = min(pt2d[1, :])
x_max = max(pt2d[0, :])
y_max = max(pt2d[1, :])
# k = 0.2 to 0.40
k = np.random.random_sample() * 0.2 + 0.2
x_min -= 0.6 * k * abs(x_max - x_min)
y_min -= 2 * k * abs(y_max - y_min)
x_max += 0.6 * k * abs(x_max - x_min)
y_max += 0.6 * k * abs(y_max - y_min)
img = img.crop((int(x_min), int(y_min), int(x_max), int(y_max)))
# We get the pose in radians
pose = utils.get_ypr_from_mat(mat_path)
pitch = pose[0] * 180 / np.pi
yaw = pose[1] * 180 / np.pi
roll = pose[2] * 180 / np.pi
ds = 1 + np.random.randint(0, 4) * 5
original_size = img.size
img = img.resize((img.size[0] / ds, img.size[1] / ds),
resample=Image.NEAREST)
img = img.resize((original_size[0], original_size[1]),
resample=Image.NEAREST)
# Flip?
rnd = np.random.random_sample()
if rnd < 0.5:
yaw = -yaw
roll = -roll
img = img.transpose(Image.FLIP_LEFT_RIGHT)
# Blur?
rnd = np.random.random_sample()
if rnd < 0.05:
img = img.filter(ImageFilter.BLUR)
# Bin values
bins = np.array(range(-99, 102, 3))
binned_pose = np.digitize([yaw, pitch, roll], bins) - 1
# Get target tensors
labels = binned_pose
cont_labels = torch.FloatTensor([yaw, pitch, roll])
if self.transform is not None:
img = self.transform(img)
return img, labels, cont_labels, self.X_train[index]
def __getitem__(self, index):
img = Image.open(
os.path.join(self.data_dir, self.X_train[index] + self.img_ext))
img = img.convert(self.image_mode)
mat_path = os.path.join(self.data_dir,
self.y_train[index] + self.annot_ext)
# Crop the face loosely
pt2d = utils.get_pt2d_from_mat(mat_path)
x_min = min(pt2d[0, :])
y_min = min(pt2d[1, :])
x_max = max(pt2d[0, :])
y_max = max(pt2d[1, :])
# k = 0.2 to 0.40
k = np.random.random_sample() * 0.2 + 0.2
x_min -= 0.6 * k * abs(x_max - x_min)
y_min -= 2 * k * abs(y_max - y_min)
x_max += 0.6 * k * abs(x_max - x_min)
y_max += 0.6 * k * abs(y_max - y_min)
img = img.crop((int(x_min), int(y_min), int(x_max), int(y_max)))
# We get the pose in radians
pose = utils.get_ypr_from_mat(mat_path)
# And convert to degrees.
# if any(pose*180/np.pi <= -99) or any(pose*180/np.pi >= 99):
# with open(os.path.join(self.data_dir, "image_garbage.txt"), 'a') as file:
# file.write(self.X_train[index] + '\n')
# file.close()
pitch = min(max(pose[0] * 180 / np.pi, -98.99), 98.99)
yaw = min(max(pose[1] * 180 / np.pi, -98.99), 98.99)
roll = min(max(pose[2] * 180 / np.pi, -98.99), 98.99)
# Flip?
rnd = np.random.random_sample()
if rnd < 0.5:
yaw = -yaw
roll = -roll
img = img.transpose(Image.FLIP_LEFT_RIGHT)
# Blur?
rnd = np.random.random_sample()
if rnd < 0.05:
img = img.filter(ImageFilter.BLUR)
# Bin values
bins = np.array(range(-99, 102, 3))
binned_pose = np.digitize([yaw, pitch, roll], bins) - 1
# Get target tensors
labels = binned_pose
cont_labels = torch.FloatTensor([yaw, pitch, roll])
if any(labels < 0) or any(labels >= 66):
print("Out labels error {}, yxz: {}, labels: {}".format(
self.X_train[index], [yaw, pitch, roll], labels))
if self.transform is not None:
img = self.transform(img)
return img, labels, cont_labels, self.X_train[index]
def __getitem__(self, index):
img = Image.open(
os.path.join(self.data_dir, self.X_train[index] + self.img_ext))
img = img.convert(self.image_mode)
mat_path = os.path.join(self.data_dir,
self.y_train[index] + self.annot_ext)
# Crop the face loosely
pt2d = utils.get_pt2d_from_mat(mat_path)
x_min = min(pt2d[0, :])
y_min = min(pt2d[1, :])
x_max = max(pt2d[0, :])
y_max = max(pt2d[1, :])
# k = 0.2 to 0.40
k = np.random.random_sample() * 0.2 + 0.2
x_min -= 0.6 * k * abs(x_max - x_min)
y_min -= 2 * k * abs(y_max - y_min)
x_max += 0.6 * k * abs(x_max - x_min)
y_max += 0.6 * k * abs(y_max - y_min)
img = img.crop((int(x_min), int(y_min), int(x_max), int(y_max)))
# We get the pose in radians
pose = utils.get_ypr_from_mat(mat_path)
# And convert to degrees.
pitch = pose[0] * 180 / np.pi
yaw = pose[1] * 180 / np.pi
roll = pose[2] * 180 / np.pi
# Flip?
rnd = np.random.random_sample()
if rnd < 0.5:
yaw = -yaw
roll = -roll
img = img.transpose(Image.FLIP_LEFT_RIGHT)
# Blur?
rnd = np.random.random_sample()
if rnd < 0.05:
img = img.filter(ImageFilter.BLUR)
# Bin values
bins = np.array(range(-99, 100, 3))
binned_pose = np.digitize([yaw, pitch, roll], bins) - 1
map_to_minus = np.where(binned_pose == 66)
binned_pose[map_to_minus] = -1
# Get target tensors
labels = binned_pose
cont_labels = torch.tensor([yaw, pitch, roll],
dtype=torch.float,
requires_grad=False)
if self.transform is not None:
img = self.transform(img)
return img, labels, cont_labels, self.X_train[index]
def __getitem__(self, index):
img = Image.open(os.path.join(self.data_dir, self.X_train[index] + self.img_ext)) # 读取一张图像
im = cv2.imread(os.path.join(self.data_dir, self.X_train[index] + self.img_ext)) # 读取一张图像
img = img.convert(self.image_mode)
mat_path = os.path.join(self.data_dir, self.y_train[index] + self.annot_ext) # 读取该图像的mat格式标签
# Crop the face loosely
pt2d = utils.get_pt2d_from_mat(mat_path)
x_min = min(pt2d[0,:])
y_min = min(pt2d[1,:])
x_max = max(pt2d[0,:])
y_max = max(pt2d[1,:])
# k = 0.2 to 0.40
k = np.random.random_sample() * 0.2 + 0.2
x_min -= 0.6 * k * abs(x_max - x_min)
y_min -= 2 * k * abs(y_max - y_min)
x_max += 0.6 * k * abs(x_max - x_min)
y_max += 0.6 * k * abs(y_max - y_min)
img = img.crop((int(x_min), int(y_min), int(x_max), int(y_max)))
im = im[int(y_min):int(y_max), int(x_min):int(x_max)]
#cv2.imwrite('1.jpg', im) # check whether the labels is correct
# We get the pose in radians
pose = utils.get_ypr_from_mat(mat_path)
# And convert to degrees.
pitch = pose[0] * 180 / np.pi
yaw = pose[1] * 180 / np.pi
roll = pose[2] * 180 / np.pi
# Flip?
rnd = np.random.random_sample()
if rnd < 0.5:
yaw = -yaw
roll = -roll
img = img.transpose(Image.FLIP_LEFT_RIGHT)
# Blur?
rnd = np.random.random_sample()
if rnd < 0.05:
img = img.filter(ImageFilter.BLUR)
# Bin values
bins = np.array(range(-99, 102, 3))
binned_pose = np.digitize([yaw, pitch, roll], bins) # - 1
# Get target tensors
labels = binned_pose
cont_labels = torch.FloatTensor([yaw, pitch, roll])
if self.transform is not None:
img = self.transform(img)
return img, labels, cont_labels, self.X_train[index]
def __getitem__(self, index):
img = Image.open(os.path.join(self.data_dir, self.X_train[index].split('.')[0] + self.img_ext))
img = img.convert(self.image_mode)
mat_path = os.path.join(self.data_dir, self.y_train[index].split('.')[0] + self.annot_ext)
# Crop the face loosely
pt2d = utils.get_pt2d_from_mat(mat_path)
x_min = min(pt2d[0,:])
y_min = min(pt2d[1,:])
x_max = max(pt2d[0,:])
y_max = max(pt2d[1,:])
# k = 0.2 to 0.40
k = np.random.random_sample() * 0.2 + 0.2
x_min -= 0.6 * k * abs(x_max - x_min)
y_min -= 2 * k * abs(y_max - y_min)
x_max += 0.6 * k * abs(x_max - x_min)
y_max += 0.6 * k * abs(y_max - y_min)
img = img.crop((int(x_min), int(y_min), int(x_max), int(y_max)))
# We get the pose in radians
pose = utils.get_ypr_from_mat(mat_path)
# And convert to degrees.
pitch = pose[0] * 180 / np.pi
yaw = pose[1] * 180 / np.pi
roll = pose[2] * 180 / np.pi
# Flip?
rnd = np.random.random_sample()
if rnd < 0.5:
yaw = -yaw
roll = -roll
img = img.transpose(Image.FLIP_LEFT_RIGHT)
# Blur?
rnd = np.random.random_sample()
if rnd < 0.05:
img = img.filter(ImageFilter.BLUR)
# Bin values: 200 bins
# 200 bins - Bin width 1 ; 40 bins - bin width 5; 66 bins - Bin Width 3
bin_width = int((102 - (-99))/self.num_bins)
bins = np.array(range(-99, 102, bin_width))
binned_pose = np.digitize([yaw, pitch, roll], bins) - 1
# Get target tensors
labels = binned_pose
cont_labels = torch.FloatTensor([yaw, pitch, roll])
if self.transform is not None:
img = self.transform(img)
return img, labels, cont_labels, self.X_train[index]
def __getitem__(self, index):
img = Image.open(
os.path.join(self.data_dir, self.X_train[index] + self.img_ext))
img = img.convert(self.image_mode)
mat_path = os.path.join(self.data_dir,
self.y_train[index] + self.annot_ext)
# Crop the face loosely
pt2d = utils.get_pt2d_from_mat(mat_path)
x_min = min(pt2d[0, :])
y_min = min(pt2d[1, :])
x_max = max(pt2d[0, :])
y_max = max(pt2d[1, :])
k = 0.20
x_min -= 2 * k * abs(x_max - x_min)
y_min -= 2 * k * abs(y_max - y_min)
x_max += 2 * k * abs(x_max - x_min)
y_max += 0.6 * k * abs(y_max - y_min)
img = img.crop((int(x_min), int(y_min), int(x_max), int(y_max)))
ds = 3 # downsampling factor
original_size = img.size
img = img.resize((img.size[0] / ds, img.size[1] / ds),
resample=Image.NEAREST)
img = img.resize((original_size[0], original_size[1]),
resample=Image.NEAREST)
# We get the pose in radians
pose = utils.get_ypr_from_mat(mat_path)
# And convert to degrees.
pitch = pose[0] * 180 / np.pi
yaw = pose[1] * 180 / np.pi
roll = pose[2] * 180 / np.pi
# Bin values
bins = np.array(range(-99, 102, 3))
binned_pose = np.digitize([yaw, pitch, roll], bins) - 1
map_to_minus = np.where(binned_pose == 66)
binned_pose[map_to_minus] = -1
labels = torch.LongTensor(binned_pose)
cont_labels = torch.FloatTensor([yaw, pitch, roll])
if self.transform is not None:
img = self.transform(img)
return img, labels, cont_labels, self.X_train[index]
def generate(self):
for index in range(self.length):
img = Image.open(os.path.join(self.data_dir, self.X_train[index] + self.img_ext))
img = img.convert(self.image_mode)
mat_path = os.path.join(self.data_dir, self.y_train[index] + self.annot_ext)
# Crop the face loosely
pt2d = utils.get_pt2d_from_mat(mat_path)
x_min = min(pt2d[0,:])
y_min = min(pt2d[1,:])
x_max = max(pt2d[0,:])
y_max = max(pt2d[1,:])
# k = 0.2 to 0.40
k = np.random.random_sample() * 0.2 + 0.2
x_min -= 0.6 * k * abs(x_max - x_min)
y_min -= 2 * k * abs(y_max - y_min)
x_max += 0.6 * k * abs(x_max - x_min)
y_max += 0.6 * k * abs(y_max - y_min)
img = img.crop((int(x_min), int(y_min), int(x_max), int(y_max)))
# We get the pose in radians
pose = utils.get_ypr_from_mat(mat_path)
# And convert to degrees.
pitch = pose[0] * 180 / np.pi
yaw = pose[1] * 180 / np.pi
roll = pose[2] * 180 / np.pi
# Bin values
bins = np.array(range(-99, 99, 3))
binned_pose = np.digitize([yaw, pitch, roll], bins)
# Get target tensors
labels = binned_pose
cont_labels = np.array([yaw, pitch, roll])
yield img, labels, cont_labels, self.X_train[index]
def __getitem__(self, index):
wdata = open("/media/omnisky/D4T/huli/work/headpose/data/filename_mix_biwi_300e_lp.txt",'a')
img_path = os.path.join(self.data_dir, self.X_train[index] + self.img_ext)
# img = Image.open(os.path.join(self.data_dir, self.X_train[index] + self.img_ext))
# img = img.convert(self.image_mode)
mat_path = os.path.join(self.data_dir, self.y_train[index] + self.annot_ext)
# Crop the face loosely
pt2d = utils.get_pt2d_from_mat(mat_path)
x_min = min(pt2d[0,:])
y_min = min(pt2d[1,:])
x_max = max(pt2d[0,:])
y_max = max(pt2d[1,:])
# k = 0.35 to 0.650
k = np.random.random_sample() * 0.35 + 0.3
w,h = x_max - x_min,y_max - y_min
ratio = h/w - 1
x_min -= (ratio/2*w+k*h)
y_min -= (k*h+40)
# x_min -= 0.6 * k * abs(x_max - x_min)
# y_min -= 2 * k * abs(y_max - y_min)
x_max += (ratio/2*w+k*h)
y_max += (k*h-40)
# x_min -= 0.6 * k * abs(x_max - x_min)
# y_min -= 2 * k * abs(y_max - y_min)
# x_max += 0.6 * k * abs(x_max - x_min)
# y_max += 0.6 * k * abs(y_max - y_min)
# img = img.crop((int(x_min), int(y_min), int(x_max), int(y_max)))
# We get the pose in radians
pose = utils.get_ypr_from_mat(mat_path)
# And convert to degrees.
pitch = pose[0] * 180 / np.pi
yaw = pose[1] * 180 / np.pi
roll = pose[2] * 180 / np.pi
# Flip?
# rnd = np.random.random_sample()
# if rnd < 0.5:
# yaw = -yaw
# roll = -roll
# img = img.transpose(Image.FLIP_LEFT_RIGHT)
# Blur?
# rnd = np.random.random_sample()
# if rnd < 0.05:
# img = img.filter(ImageFilter.BLUR)
# erase
# img = RandomErasing()(img)
# Bin values
# bins = np.array(range(-99, 102, 3))
# binned_pose = np.digitize([yaw, pitch, roll], bins) - 1
# Get target tensors
# labels = binned_pose
# cont_labels = torch.FloatTensor([yaw, pitch, roll])
# if self.transform is not None:
# img = self.transform(img)
wdata.write("{},{},{},{},{},{},{},{}\n".format(img_path,x_min,y_min,x_max,y_max,yaw,pitch,roll))
wdata.close()
return 0
def __getitem__(self, index):
img = Image.open(os.path.join(self.data_dir, self.X_train[index] + self.img_ext))
img = img.convert(self.image_mode)
mat_path = os.path.join(self.data_dir, self.y_train[index] + self.annot_ext)
# Crop the face loosely
pt2d = utils.get_pt2d_from_mat(mat_path)
x_min = min(pt2d[0,:])
y_min = min(pt2d[1,:])
x_max = max(pt2d[0,:])
y_max = max(pt2d[1,:])
# k = 0.2 to 0.40
k = np.random.random_sample() * 0.2 + 0.2
x_min -= 0.6 * k * abs(x_max - x_min)
y_min -= 2 * k * abs(y_max - y_min)
x_max += 0.6 * k * abs(x_max - x_min)
y_max += 0.6 * k * abs(y_max - y_min)
img = img.crop((int(x_min), int(y_min), int(x_max), int(y_max)))
# We get the pose in radians
pose = utils.get_ypr_from_mat(mat_path)
pitch = pose[0] * 180 / np.pi
yaw = pose[1] * 180 / np.pi
roll = pose[2] * 180 / np.pi
ds = 1 + np.random.randint(0,4) * 5
original_size = img.size
img = img.resize((img.size[0] // ds, img.size[1] // ds), resample=Image.NEAREST)
img = img.resize((original_size[0], original_size[1]), resample=Image.NEAREST)
# Rotate
angle = np.random.uniform(-50, 50)
R = utils.create_rotation_matrix(yaw, pitch, roll)
R = np.matmul(R, utils.create_rotation_matrix(0, 0, -angle))
yaw_r, pitch_r, roll_r = utils.rotation_matrix_to_euler_angles(R)
if np.abs(yaw_r) <= self.max_angle and np.abs(pitch_r) <= self.max_angle and np.abs(roll_r) <= self.max_angle:
img = img.rotate(angle, resample=Image.NEAREST)
#utils.draw_axis_pil(img, yaw_r, pitch_r, roll_r).show()
yaw, pitch, roll = yaw_r, pitch_r, roll_r
#exit()
# Flip?
rnd = np.random.random_sample()
if rnd < 0.5:
yaw = -yaw
roll = -roll
img = img.transpose(Image.FLIP_LEFT_RIGHT)
# Change contrast?
rnd = np.random.random_sample()
if rnd < 0.35:
if np.random.random_sample() < 0.5:
img = ImageEnhance.Contrast(img).enhance(np.random.uniform(0.1, 0.9))
else:
img = ImageEnhance.Contrast(img).enhance(np.random.uniform(1.1, 1.9))
# Change brightness?
rnd = np.random.random_sample()
if rnd < 0.35:
if np.random.random_sample() < 0.5:
img = ImageEnhance.Brightness(img).enhance(np.random.uniform(0.15, 0.9))
else:
img = ImageEnhance.Brightness(img).enhance(np.random.uniform(1.1, 1.85))
# Change sharpness?
rnd = np.random.random_sample()
if rnd < 0.1:
img = ImageEnhance.Sharpness(img).enhance(np.random.uniform(0.1, 1.9))
# Grayscale?
rnd = np.random.random_sample()
if rnd < 0.35:
img = img.convert('L')
img = img.convert(mode='RGB')
# Blur?
rnd = np.random.random_sample()
if rnd < 0.05:
img = img.filter(ImageFilter.BLUR)
# Bin values
binned_pose = np.digitize([yaw, pitch, roll], self.bins) - 1
# Get target tensors
labels = binned_pose
cont_labels = torch.FloatTensor([yaw, pitch, roll])
if self.transform is not None:
img = self.transform(img)
return img, labels, cont_labels, self.X_train[index]
def __getitem__(self, index):
file_name, flag = self.X_train[index].split(",")
ann_name, _ = self.y_train[index].split(",")
yaw, pitch, roll = 0, 0, 0
if int(flag) == 1:
img = cv2.imread(
os.path.join(self.data_dir,
file_name + '_rgb' + self.biwi_img_ext))
img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
img = cv2.cvtColor(img, cv2.COLOR_GRAY2RGB)
img = Image.fromarray(img.astype(np.uint8))
# img = Image.open(os.path.join(self.data_dir, file_name + '_rgb' + self.biwi_img_ext))
img = img.convert(self.image_mode)
# orial_img = img.copy()
pose_path = os.path.join(self.data_dir,
ann_name + '_pose' + self.biwi_annot_ext)
y_train_list = ann_name.split('/')
y_train_new = y_train_list[0:-1]
temp = ''
for idt in y_train_new:
temp = os.path.join(temp, idt)
y_train_new = temp
bbox_path = os.path.join(
self.data_dir, y_train_new + '/dockerface-' +
y_train_list[-1] + '_rgb' + self.biwi_annot_ext)
# Load bounding box
# bbox = open(bbox_path, 'r')
with open(bbox_path, 'r') as tf:
for tdata in tf.readlines():
line = tdata.split(" ")
if float(line[1]) > 215.0:
break
# line = bbox.readline().split(' ')
if len(line) < 4:
x_min, y_min, x_max, y_max = 0, 0, img.size[0], img.size[1]
else:
x_min, y_min, x_max, y_max = [
float(line[1]),
float(line[2]),
float(line[3]),
float(line[4])
]
# bbox.close()
# Load pose in degrees
pose_annot = open(pose_path, 'r')
R = []
for line in pose_annot:
line = line.strip('\n').split(' ')
l = []
if line[0] != '':
for nb in line:
if nb == '':
continue
l.append(float(nb))
R.append(l)
R = np.array(R)
T = R[3, :]
R = R[:3, :]
pose_annot.close()
R = np.transpose(R)
roll = -np.arctan2(R[1][0], R[0][0]) * 180 / np.pi
yaw = -np.arctan2(-R[2][0],
np.sqrt(R[2][1]**2 + R[2][2]**2)) * 180 / np.pi
pitch = np.arctan2(R[2][1], R[2][2]) * 180 / np.pi
# Loosely crop face
k = np.random.random_sample() * 0.3 + 0.15
# k = 0.35
w, h = x_max - x_min, y_max - y_min
ratio = h / w - 1
x_min -= ((ratio / 2.0 * w) + k * h) #w*k*0.6 k*(y_max - y_min)
y_min -= (k * h + 10)
# y_min -= (k*abs(y_max - y_min)+0)#h*k
x_max += (ratio / 2.0 * w) + k * h #w*h*0.6 + k*(y_max - y_min)
y_max += (k * h - 10)
# x_min -= 0.6 * k * abs(x_max - x_min)
# y_min -= k * abs(y_max - y_min)
# x_max += 0.6 * k * abs(x_max - x_min)
# y_max += 0.6 * k * abs(y_max - y_min)
img = img.crop((int(x_min), int(y_min), int(x_max), int(y_max)))
# Bin values
bins = np.array(range(-99, 102, 3))
binned_pose = np.digitize([yaw, pitch, roll], bins) - 1
# Flip?
# rnd = np.random.random_sample()
# if rnd < 0.5:
# yaw = -yaw
# roll = -roll
# img = img.transpose(Image.FLIP_LEFT_RIGHT)
rnd = np.random.random_sample()
if rnd < 0.5:
yaw = -yaw
roll = -roll
img = img.transpose(Image.FLIP_LEFT_RIGHT)
# Blur?
rnd = np.random.random_sample()
if rnd < 0.05:
img = img.filter(ImageFilter.BLUR)
# orial_img = img.copy()
# erase
# img = RandomErasing()(img)
if self.transform is not None:
img = self.transform(img)
labels = torch.LongTensor(binned_pose)
cont_labels = torch.FloatTensor([yaw, pitch, roll])
# file_name = torch.tensor(file_name)
# return img, labels, cont_labels, file_name
else:
img = cv2.imread(
os.path.join(self.data_dir, file_name + self.t300w_lp_img_ext))
img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
img = cv2.cvtColor(img, cv2.COLOR_GRAY2RGB)
img = Image.fromarray(img.astype(np.uint8))
# img = Image.open(os.path.join(self.data_dir, file_name + '_rgb' + self.biwi_img_ext))
# img = img.convert(self.image_mode)
# img = Image.open(os.path.join(self.data_dir, file_name + self.t300w_lp_img_ext))
img = img.convert(self.image_mode)
mat_path = os.path.join(self.data_dir,
ann_name + self.t300w_lp_annot_ext)
# Crop the face loosely
pt2d = utils.get_pt2d_from_mat(mat_path)
x_min = min(pt2d[0, :])
y_min = min(pt2d[1, :])
x_max = max(pt2d[0, :])
y_max = max(pt2d[1, :])
# k = 0.35 to 0.650
k = np.random.random_sample() * 0.35 + 0.3
w, h = x_max - x_min, y_max - y_min
ratio = h / w - 1
x_min -= (ratio / 2 * w + k * h)
y_min -= (k * h + 40)
# x_min -= 0.6 * k * abs(x_max - x_min)
# y_min -= 2 * k * abs(y_max - y_min)
x_max += (ratio / 2 * w + k * h)
y_max += (k * h - 40)
# x_min -= 0.6 * k * abs(x_max - x_min)
# y_min -= 2 * k * abs(y_max - y_min)
# x_max += 0.6 * k * abs(x_max - x_min)
# y_max += 0.6 * k * abs(y_max - y_min)
img = img.crop((int(x_min), int(y_min), int(x_max), int(y_max)))
# We get the pose in radians
pose = utils.get_ypr_from_mat(mat_path)
# And convert to degrees.
pitch = pose[0] * 180 / np.pi
yaw = pose[1] * 180 / np.pi
roll = pose[2] * 180 / np.pi
# Flip?
rnd = np.random.random_sample()
if rnd < 0.5:
yaw = -yaw
roll = -roll
img = img.transpose(Image.FLIP_LEFT_RIGHT)
# Blur?
rnd = np.random.random_sample()
if rnd < 0.05:
img = img.filter(ImageFilter.BLUR)
# erase
# img = RandomErasing()(img)
# orial_img = img.copy()
# Bin values
bins = np.array(range(-99, 102, 3))
binned_pose = np.digitize([yaw, pitch, roll], bins) - 1
# Get target tensors
labels = torch.LongTensor(binned_pose)
cont_labels = torch.FloatTensor([yaw, pitch, roll])
if self.transform is not None:
img = self.transform(img)
# print(file_name)
file_name = file_name.split('/')[-1]
return img, labels, cont_labels, file_name
def __getitem__(self, index):
img = cv2.imread(
os.path.join(self.data_dir, self.X_train[index] + self.img_ext))
img = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
img = cv2.cvtColor(img, cv2.COLOR_GRAY2RGB)
img = Image.fromarray(img.astype(np.uint8))
# img = Image.open(os.path.join(self.data_dir, self.X_train[index] + self.img_ext))
img = img.convert(self.image_mode)
mat_path = os.path.join(self.data_dir,
self.y_train[index] + self.annot_ext)
# Crop the face loosely
pt2d = utils.get_pt2d_from_mat(mat_path)
ih, iw = img.height, img.width
x_list = pt2d[0, :]
x_list = np.sort(x_list)
r_x_list = x_list[np.argsort(-x_list)]
for i in range(len(x_list)):
if x_list[i] != -1:
x_min = x_list[i]
break
for i in range(len(x_list)):
if r_x_list[i] < iw:
x_max = r_x_list[i]
break
y_list = pt2d[1, :]
y_list = np.sort(y_list)
r_y_list = y_list[np.argsort(-y_list)]
for i in range(len(y_list)):
if y_list[i] != -1:
y_min = y_list[i]
break
for i in range(len(y_list)):
if ih > r_y_list[i]:
y_max = r_y_list[i]
break
w, h = x_max - x_min, y_max - y_min
ratio = h / w
k = 0.4
if ratio > 1:
x_min = max(x_min - w * (ratio - 1.0) / 2.0 - h * k, 0)
x_max = min(x_max + w * (ratio - 1.0) / 2.0 + h * k, iw)
y_min = max(y_min - h * k - 35.0, 0)
y_max = min(y_max + h * k - 35.0, ih)
else:
ratio = w / h
y_min = max(y_min - h * (ratio - 1.0) / 2.0 - w * k - 35.0, 0)
y_max = min(y_max + h * (ratio - 1.0) / 2.0 + w * k - 35.0, ih)
x_min = max(x_min - w * k, 0)
x_max = min(x_max + w * k, iw)
# x_min = min(pt2d[0, :])
# y_min = min(pt2d[1, :])
# x_max = max(pt2d[0, :])
# y_max = max(pt2d[1, :])
# k = 0.20
# x_min -= 2 * k * abs(x_max - x_min)
# y_min -= 2 * k * abs(y_max - y_min)
# x_max += 2 * k * abs(x_max - x_min)
# y_max += 0.6 * k * abs(y_max - y_min)
img = img.crop((int(x_min), int(y_min), int(x_max), int(y_max)))
# We get the pose in radians
pose = utils.get_ypr_from_mat(mat_path)
# And convert to degrees.
pitch = pose[0] * 180 / np.pi
yaw = pose[1] * 180 / np.pi
roll = pose[2] * 180 / np.pi
# Bin values
bins = np.array(range(-99, 102, 3))
labels = torch.LongTensor(np.digitize([yaw, pitch, roll], bins) - 1)
cont_labels = torch.FloatTensor([yaw, pitch, roll])
if self.transform is not None:
img = self.transform(img)
return img, labels, cont_labels, self.X_train[index]
def __getitem__(self, index):
img = Image.open(
os.path.join(self.data_dir, self.X_train[index] + self.img_ext))
img = img.convert(self.image_mode)
mat_path = os.path.join(self.data_dir,
self.y_train[index] + self.annot_ext)
# Crop the face loosely
pt2d = utils.get_pt2d_from_mat(mat_path)
x_min = min(pt2d[0, :])
y_min = min(pt2d[1, :])
x_max = max(pt2d[0, :])
y_max = max(pt2d[1, :])
# k = 0.35 to 0.650
k = np.random.random_sample() * 0.35 + 0.3
w, h = x_max - x_min, y_max - y_min
ratio = h / w - 1
x_min -= (ratio / 2 * w + k * h)
y_min -= (k * h + 40)
# x_min -= 0.6 * k * abs(x_max - x_min)
# y_min -= 2 * k * abs(y_max - y_min)
x_max += (ratio / 2 * w + k * h)
y_max += (k * h - 40)
# x_min -= 0.6 * k * abs(x_max - x_min)
# y_min -= 2 * k * abs(y_max - y_min)
# x_max += 0.6 * k * abs(x_max - x_min)
# y_max += 0.6 * k * abs(y_max - y_min)
img = img.crop((int(x_min), int(y_min), int(x_max), int(y_max)))
# We get the pose in radians
pose = utils.get_ypr_from_mat(mat_path)
# And convert to degrees.
pitch = pose[0] * 180 / np.pi
yaw = pose[1] * 180 / np.pi
roll = pose[2] * 180 / np.pi
# Flip?
rnd = np.random.random_sample()
if rnd < 0.5:
yaw = -yaw
roll = -roll
img = img.transpose(Image.FLIP_LEFT_RIGHT)
# Blur?
rnd = np.random.random_sample()
if rnd < 0.05:
img = img.filter(ImageFilter.BLUR)
# erase
img = RandomErasing()(img)
# Bin values
bins = np.array(range(-99, 102, 3))
binned_pose = np.digitize([yaw, pitch, roll], bins) - 1
# Get target tensors
labels = binned_pose
cont_labels = torch.FloatTensor([yaw, pitch, roll])
if self.transform is not None:
img = self.transform(img)
return img, labels, cont_labels, self.X_train[index]
def get(self):
images = np.zeros(
(self.batch_size, self.image_size, self.image_size, 3))
Llabels = np.zeros((self.batch_size, 3), np.int32)
Lcont_labels = np.zeros((self.batch_size, 3))
count = 0
while count < self.batch_size:
img = Image.open(
os.path.join(self.data_dir,
self.X_train[self.cursor] + self.img_ext))
#print('img', img.shape)
img = img.convert(self.image_mode)
mat_path = os.path.join(self.data_dir,
self.y_train[self.cursor] + self.annot_ext)
# Crop the face loosely
pt2d = utils.get_pt2d_from_mat(mat_path)
x_min = min(pt2d[0, :])
y_min = min(pt2d[1, :])
x_max = max(pt2d[0, :])
y_max = max(pt2d[1, :])
# k = 0.2 to 0.40
k = np.random.random_sample() * 0.2 + 0.2
x_min -= 0.6 * k * abs(x_max - x_min)
y_min -= 2 * k * abs(y_max - y_min)
x_max += 0.6 * k * abs(x_max - x_min)
y_max += 0.6 * k * abs(y_max - y_min)
img = img.crop((int(x_min), int(y_min), int(x_max), int(y_max)))
# We get the pose in radians
pose = utils.get_ypr_from_mat(mat_path)
pitch = pose[0] * 180 / np.pi
yaw = pose[1] * 180 / np.pi
roll = pose[2] * 180 / np.pi
# Flip?
rnd = np.random.random_sample()
if rnd < 0.5:
yaw = -yaw
roll = -roll
img = img.transpose(Image.FLIP_LEFT_RIGHT)
# Blur?
rnd = np.random.random_sample()
if rnd < 0.05:
img = img.filter(ImageFilter.BLUR)
#preprocess
img = rescale(img)
img = random_crop(img)
img = nomalizing(img, [0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
# Bin values
bins = np.array(range(-99, 102, 3))
binned_pose = np.digitize([yaw, pitch, roll], bins) - 1
labels = binned_pose
cont_labels = [float(yaw), float(pitch), float(roll)]
images[count, :, :, :] = img
Llabels[count] = labels
Lcont_labels[count] = cont_labels
count += 1
self.cursor += 1
if self.cursor >= len(self.X_train):
np.random.shuffle(self.X_train)
self.cursor = 0
print("self.cursor ====0")
#print(self.X_train[0])
return images, Llabels, Lcont_labels