def valid_generator(sample_per_batch, batch_number):
""" Generating validation data """
valid_image_file = []
directory = '../EgoGesture Dataset/'
folder_name = [
'SingleOneValid', 'SingleTwoValid', 'SingleThreeValid',
'SingleFourValid', 'SingleFiveValid', 'SingleSixValid',
'SingleSevenValid', 'SingleEightValid'
]
for folder in folder_name:
valid_image_file = valid_image_file + os.listdir(directory + folder +
'/')
# print(len(valid_image_file))
while True:
for i in range(0, batch_number - 1):
start = i * sample_per_batch
end = (i + 1) * sample_per_batch
x_batch = []
y_batch_prob = []
y_batch_pos = []
for n in range(start, end):
image_name = valid_image_file[n]
try:
image, probability, position = label_generator(
directory=directory,
image_name=image_name,
type='Valid')
except cv2.error:
print(image_name)
continue
# 1.0 Original image
x_batch.append(image)
y_batch_prob.append(probability)
pos = np.array([
position,
] * 10)
y_batch_pos.append(pos)
x_batch = np.asarray(x_batch)
x_batch = x_batch.astype('float32')
x_batch = x_batch / 255.
y_batch_prob = np.asarray(y_batch_prob)
y_batch_pos = np.asarray(y_batch_pos)
y_batch_pos = y_batch_pos.astype('float32')
y_batch_pos = y_batch_pos / 128.
y_batch = [y_batch_prob, y_batch_pos]
yield (x_batch, y_batch)
def valid_generator(sample_per_batch, batch_number):
""" Generating validation data """
valid_image_files = os.listdir(valid_directory)
for i in range(0, 10):
random.shuffle(valid_image_files)
while True:
for i in range(0, batch_number - 1):
start = i * sample_per_batch
end = (i + 1) * sample_per_batch
x_batch = []
y_batch_key = []
for n in range(start, end):
image_name = valid_image_files[n]
# image: ndarray keypoints: ndarray but not normalized
if 'TI1K' in image_name.split('_'):
image, keypoints = my_label_gen(
image_directory=valid_directory,
label_directory=label_directory,
image_name=image_name)
else:
image, keypoints = label_generator(
image_directory=valid_directory,
label_directory=label_directory,
image_name=image_name)
x_batch.append(image)
y_batch_key.append(keypoints)
im_flip, k_flip = flip_horizontal(image, keypoints)
x_batch.append(im_flip)
y_batch_key.append(k_flip)
x_batch = np.asarray(x_batch)
x_batch = x_batch.astype('float32')
x_batch = x_batch / 255.
y_batch_key = np.asarray(y_batch_key)
y_batch_key = y_batch_key.astype('float32')
y_batch_key = y_batch_key / 128.
yield (x_batch, y_batch_key)
]
index = 0
for i in range(0, len(prob)):
output[index] = output[index] * prob[i]
output[index + 1] = output[index + 1] * prob[i]
index = index + 2
output = np.array(output)
return image_aug, output
if __name__ == '__main__':
directory = '../../EgoGesture Dataset/'
image_name = 'ChuangyeguBusstop_Single_Five_color_1.jpg'
image, probability, position = label_generator(directory=directory,
image_name=image_name,
type='')
""" augmentation """
image, position = rotation(image, probability, position)
# image, position = translate(image, probability, position)
# image, position = crop(image, probability, position)
# image, position = flip_horizontal(image, probability, position)
# image, position = flip_vertical(image, probability, position)
# image, position = noise(image, probability, position)
# image, position = salt(image, probability, position)
print(probability)
print(position)
# draw augmented keypoints
index = 0
color = [(120, 20, 240), (240, 55, 210), (240, 55, 140), (240, 75, 55),
keys_aug = np.asarray(keys_aug)
return image_aug, keys_aug
if __name__ == '__main__':
folder_directory = '../../Dataset/Train/'
label_directory = '../../Dataset/label/'
image_name = 'EastLake_Single_Eight_color_4.jpg'
if 'TI1K' in image_name.split('_'):
image, keys = my_label_gen(image_directory=folder_directory,
label_directory=label_directory,
image_name=image_name)
else:
image, keys = label_generator(image_directory=folder_directory,
label_directory=label_directory,
image_name=image_name)
image, k = rotation(image, keys)
# image, k = translate(image, keys)
# image, k = crop(image, keys)
# image, k = flip_horizontal(image, keys)
# image, k = flip_vertical(image, keys)
# image, k = noise(image, keys)
# image, k = salt(image, keys)
image = cv2.circle(image, (int(k[0]), int(k[1])), 5, (0, 0, 255), -1)
image = cv2.circle(image, (int(k[2]), int(k[3])), 5, (0, 255, 0), -1)
cv2.imshow('', image)
while True:
import cv2
from preprocess.datagen import label_generator, my_label_gen
def visualize(image, keys):
""" visualizing the image and the output data """
image = cv2.circle(image, (int(keys[0]), int(keys[1])), 5, (0, 0, 255), -1)
image = cv2.circle(image, (int(keys[2]), int(keys[3])), 5, (0, 255, 0), -1)
cv2.imshow('visual output', image)
cv2.waitKey(0)
if __name__ == '__main__':
image_directory = '../Dataset/Train/'
label_directory = '../Dataset/label/'
image_name = 'TI1K_IMAGE_0212.jpg'
if 'TI1K' in image_name.split('_'):
image, keypoints = my_label_gen(image_directory=image_directory,
label_directory=label_directory,
image_name=image_name)
else:
image, keypoints = label_generator(image_directory=image_directory,
label_directory=label_directory,
image_name=image_name)
visualize(image, keypoints)
def train_generator(sample_per_batch, batch_number):
""" Generating training data """
train_image_file = []
directory = '../EgoGesture Dataset/'
folder_name = [
'SingleOne', 'SingleTwo', 'SingleThree', 'SingleFour', 'SingleFive',
'SingleSix', 'SingleSeven', 'SingleEight'
]
for folder in folder_name:
train_image_file = train_image_file + os.listdir(directory + folder +
'/')
for i in range(0, 10):
random.shuffle(train_image_file)
print('Training Dataset Size: {0}'.format(len(train_image_file)))
while True:
for i in range(0, batch_number - 1):
start = i * sample_per_batch
end = (i + 1) * sample_per_batch
x_batch = []
y_batch_prob = []
y_batch_pos = []
for n in range(start, end):
image_name = train_image_file[n]
try:
image, probability, position = label_generator(
directory=directory, image_name=image_name, type='')
except cv2.error:
print(image_name)
continue
# 1.0 Original image
x_batch.append(image)
y_batch_prob.append(probability)
pos = np.array([
position,
] * 10)
y_batch_pos.append(pos)
# visualize(image, probability, pos[0])
""" Augmentation """
# 2.0 Original + translate
im, pos = translate(image, probability, position)
x_batch.append(im)
y_batch_prob.append(probability)
pos = np.array([
pos,
] * 10)
y_batch_pos.append(pos)
# visualize(im, probability, pos[0])
# 3.0 Original + rotation
im, pos = rotation(image, probability, position)
x_batch.append(im)
y_batch_prob.append(probability)
pos = np.array([
pos,
] * 10)
y_batch_pos.append(pos)
# visualize(im, probability, pos[0])
# 4.0 Original + salt
im, pos = salt(image, probability, position)
x_batch.append(im)
y_batch_prob.append(probability)
pos = np.array([
pos,
] * 10)
y_batch_pos.append(pos)
# visualize(im, probability, pos[0])
# 5.0 Original + crop
im, pos = crop(image, probability, position)
x_batch.append(im)
y_batch_prob.append(probability)
pos = np.array([
pos,
] * 10)
y_batch_pos.append(pos)
# visualize(im, probability, pos)
# 6.0 Original + noise
im, pos = noise(image, probability, position)
x_batch.append(im)
y_batch_prob.append(probability)
pos = np.array([
pos,
] * 10)
y_batch_pos.append(pos)
# visualize(im, probability, pos[0])
# 7.0 Original + rotate + translate
im, pos = rotation(image, probability, position)
im, pos = translate(im, probability, pos)
x_batch.append(im)
y_batch_prob.append(probability)
pos = np.array([
pos,
] * 10)
y_batch_pos.append(pos)
# visualize(im, probability, pos)
# 8.0 Original + rotate + crop
im, pos = rotation(image, probability, position)
im, pos = crop(im, probability, pos)
x_batch.append(im)
y_batch_prob.append(probability)
pos = np.array([
pos,
] * 10)
y_batch_pos.append(pos)
# visualize(im, probability, pos[0])
x_batch = np.asarray(x_batch)
x_batch = x_batch.astype('float32')
x_batch = x_batch / 255.
y_batch_prob = np.asarray(y_batch_prob)
y_batch_pos = np.asarray(y_batch_pos)
y_batch_pos = y_batch_pos.astype('float32')
y_batch_pos = y_batch_pos / 128.
y_batch = [y_batch_prob, y_batch_pos]
yield (x_batch, y_batch)
import cv2
from preprocess.datagen import label_generator
def visualize(image, prob, key):
index = 0
color = [(15, 15, 240), (15, 240, 155), (240, 155, 15), (240, 15, 155),
(240, 15, 240)]
for c, p in enumerate(prob):
if p > 0.5:
image = cv2.circle(image, (int(key[index]), int(key[index + 1])),
radius=5,
color=color[c],
thickness=-2)
index = index + 2
cv2.imshow('Visualize', image)
cv2.waitKey(0)
if __name__ == '__main__':
image_name = 'ChuangyeguFirstfloor_Single_Five_color_514.jpg'
image, probability, position = label_generator(image_name=image_name,
type='Test')
visualize(image, probability, position)
def train_generator(sample_per_batch, batch_number):
""" Generating training data """
train_image_file = []
image_files = os.listdir(train_directory)
train_image_file = train_image_file + image_files
print('Training Dataset Size: ' + str(len(train_image_file)))
for i in range(0, 10):
random.shuffle(train_image_file)
while True:
for i in range(0, batch_number - 1):
start = i * sample_per_batch
end = (i + 1) * sample_per_batch
x_batch = []
y_batch_key = []
for n in range(start, end):
image_name = train_image_file[n]
# image: ndarray keypoints: ndarray but not normalized
if 'TI1K' in image_name.split('_'):
image, keypoints = my_label_gen(
image_directory=train_directory,
label_directory=label_directory,
image_name=image_name)
else:
image, keypoints = label_generator(
image_directory=train_directory,
label_directory=label_directory,
image_name=image_name)
# 1.0 Original Image
x_batch.append(image)
y_batch_key.append(keypoints)
# visualize(image, keypoints)
""" Augmentation """
# 2.0 Flip
im_flip, k_flip = flip_horizontal(image, keypoints)
x_batch.append(im_flip)
y_batch_key.append(k_flip)
# visualize(im_flip, k_flip)
# 3.0 Original + rotation
im, k = rotation(image, keypoints)
x_batch.append(im)
y_batch_key.append(k)
# visualize(im, k)
# 4.0 Flip + rotation
im, k = rotation(im_flip, k_flip)
x_batch.append(im)
y_batch_key.append(k)
# visualize(im, k)
# 5.0 Original + translate
im, k = translate(image, keypoints)
x_batch.append(im)
y_batch_key.append(k)
# visualize(im, k)
# 6.0 Flip + translate
im, k = translate(im_flip, k_flip)
x_batch.append(im)
y_batch_key.append(k)
# visualize(im, k)
# 7.0 Original + crop
im, k = crop(image, keypoints)
x_batch.append(im)
y_batch_key.append(k)
# visualize(im, k)
# 8.0 Flip + crop
im, k = crop(im_flip, k_flip)
x_batch.append(im)
y_batch_key.append(k)
# visualize(im, k)
# 9.0 Original + noise
im, k = noise(image, keypoints)
x_batch.append(im)
y_batch_key.append(k)
# visualize(im, k)
# 10.0 Flip + noise
im, k = noise(im_flip, k_flip)
x_batch.append(im)
y_batch_key.append(k)
# visualize(im, k)
# 11.0 Original + salt
im, k = salt(image, keypoints)
x_batch.append(im)
y_batch_key.append(k)
# visualize(im, k)
# 12.0 Flip + salt
im, k = salt(im_flip, k_flip)
x_batch.append(im)
y_batch_key.append(k)
# visualize(im, k)
# 13.0 Original + flip vertical
im, k = flip_vertical(image, keypoints)
x_batch.append(im)
y_batch_key.append(k)
# visualize(im, k)
# 14.0 Flip + flip vertical
im, k = flip_vertical(im_flip, k_flip)
x_batch.append(im)
y_batch_key.append(k)
# visualize(im, k)
# 15.0 Original + rotate + translate
im, k = rotation(image, keypoints)
im, k = translate(im, k)
x_batch.append(im)
y_batch_key.append(k)
# visualize(im, k)
# 16.0 Flip + rotate + translate
im, k = rotation(im_flip, k_flip)
im, k = translate(im, k)
x_batch.append(im)
y_batch_key.append(k)
# visualize(im, k)
# 17.0 Original + noise + translate
im, k = noise(image, keypoints)
im, k = translate(im, k)
x_batch.append(im)
y_batch_key.append(k)
# visualize(im, k)
# 18.0 Flip + noise + translate
im, k = noise(im_flip, k_flip)
im, k = translate(im, k)
x_batch.append(im)
y_batch_key.append(k)
# visualize(im, k)
# 19.0 Original + crop + translate
im, k = crop(image, keypoints)
im, k = translate(im, k)
x_batch.append(im)
y_batch_key.append(k)
# visualize(im, k)
# 20.0 Flip + crop + translate
im, k = crop(im_flip, k_flip)
im, k = translate(im, k)
x_batch.append(im)
y_batch_key.append(k)
# visualize(im, k)
x_batch = np.asarray(x_batch)
x_batch = x_batch.astype('float32')
x_batch = x_batch / 255.
y_batch_key = np.asarray(y_batch_key)
y_batch_key = y_batch_key.astype('float32')
y_batch_key = y_batch_key / 128.
yield (x_batch, y_batch_key)
import cv2
from preprocess.datagen import label_generator
def visualize(image, prob, key):
index = 0
color = [(15, 15, 240), (15, 240, 155), (240, 155, 15), (240, 15, 155),
(240, 15, 240)]
for c, p in enumerate(prob):
if p > 0.5:
image = cv2.circle(image, (int(key[index]), int(key[index + 1])),
radius=5,
color=color[c],
thickness=-2)
index = index + 2
cv2.imshow('Unified Gesture & Fingertips Detection', image)
cv2.waitKey(0)
if __name__ == '__main__':
img_name = 'ChuangyeguFirstfloor_Single_Four_color_335.jpg'
dir = '../EgoGesture Dataset/'
img, prob, keys = label_generator(directory=dir,
image_name=img_name,
type='Test')
visualize(img, prob, keys)
