def siamese_process_batch_rgb(lines, img_path, train=True):
num = len(lines)
batch = np.zeros((num, 16, 112, 112, 3), dtype='float32')
btcof = np.zeros((num, 16, 112, 112, 3), dtype='float32')
labels = np.zeros(num, dtype='int')
for i in range(num):
path = lines[i].split(' ')[0]
label = lines[i].split(' ')[-1]
symbol = lines[i].split(' ')[1]
label = label.strip('\n')
label = int(label)
symbol = int(symbol) - 1
# imgs = os.listdir(img_path+path)
imgs = glob.glob(img_path + path + '/*.jpg')
if imgs == []:
imgs = glob.glob(img_path + path + '/*.png')
imgs.sort(key=str.lower)
if (symbol + 16) >= len(imgs):
continue
if train:
crop_x = random.randint(0, 10)
crop_y = random.randint(0, 10)
tran_x = random.randint(0, 10)
tran_y = random.randint(0, 10)
is_flip = random.randint(0, 1)
init_f = random.randint(0, 8)
init_f = 0
if (symbol + 16 + init_f) >= len(imgs):
init_f = 0
sc = random.uniform(0.2, 1)
for j in range(16):
img = imgs[symbol + j + init_f]
# image = cv2.imread(img_path + path + '/' + img)
image = cv2.imread(img)
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
im_of = np.load(img + '.npy')
mag, ang = cv2.cartToPolar(im_of[..., 0], im_of[..., 1])
im_of = np.dstack((im_of, mag))
# im_of *= 16
im_of = cv2.normalize(im_of, None, -1, 1, cv2.NORM_MINMAX)
# Re-scale
if doScale:
hh, ww = image.shape[0:2]
hh = int(hh * sc)
ww = int(ww * sc)
image = cv2.resize(image, (ww, hh))
im_of = cv2.resize(im_of, (ww, hh))
# image = cv2.resize(image, (171, 128))
if is_flip == 1 and doFlip:
image = cv2.flip(image, 1)
im_of = cv2.flip(im_of, 1)
if doTrans:
hh, ww = image.shape[0:2]
M = np.float32([[1, 0, tran_x], [0, 1, tran_y]])
image = cv2.warpAffine(image, M, (ww, hh))
im_of = cv2.warpAffine(im_of, M, (ww, hh))
if doNoisy:
noisy(image, 'gauss')
noisy(im_of, 'gauss')
# cv2.imshow('opticalflow',im_of[crop_x:crop_x + 112, crop_y:crop_y + 112, :])
# cv2.waitKey(1)
batch[i][j][:][:][:] = image[crop_x:crop_x + 112,
crop_y:crop_y + 112, :]
btcof[i][j][:][:][:] = im_of[crop_x:crop_x + 112,
crop_y:crop_y + 112, :]
labels[i] = label
else:
for j in range(16):
img = imgs[symbol + j]
# image = cv2.imread(img_path + path + '/' + img)
image = cv2.imread(img)
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
image = cv2.resize(image, (171, 128))
im_of = np.load(img + '.npy')
mag, ang = cv2.cartToPolar(im_of[..., 0], im_of[..., 1])
im_of = np.dstack((im_of, mag))
im_of = cv2.normalize(im_of, None, -1, 1, cv2.NORM_MINMAX)
batch[i][j][:][:][:] = image[8:120, 30:142, :]
btcof[i][j][:][:][:] = im_of[8:120, 30:142, :]
labels[i] = label
return batch, btcof, labels
def process_batch_rgb(lines, img_path, train=True):
num = len(lines)
batch = np.zeros((num, 16, 112, 112, 3), dtype='float32')
labels = np.zeros(num, dtype='int')
for i in range(num):
path = lines[i].split(' ')[0]
label = lines[i].split(' ')[-1]
symbol = lines[i].split(' ')[1]
label = label.strip('\n')
label = int(label)
symbol = int(symbol) - 1
# imgs = os.listdir(img_path+path)
imgs = glob.glob(img_path + path + '/*.jpg')
if imgs == []:
imgs = glob.glob(img_path + path + '/*.png')
imgs = [k for k in imgs if '.wav' not in k and 'flow' not in k]
imgs.sort(key=str.lower)
if (symbol + 16) >= len(imgs):
continue
if train:
crop_x = random.randint(0, 10)
crop_y = random.randint(0, 10)
tran_x = random.randint(0, 10)
tran_y = random.randint(0, 10)
is_flip = random.randint(0, 1)
init_f = random.randint(0, 8)
init_f = 0
if (symbol + 16 + init_f) >= len(imgs):
init_f = 0
sc = random.uniform(0.2, 1)
for j in range(16):
img = imgs[symbol + j + init_f]
# image = cv2.imread(img_path + path + '/' + img)
image = cv2.imread(img)
# Re-scale
if doScale:
hh, ww = image.shape[0:2]
hh = int(hh * sc)
ww = int(ww * sc)
image = cv2.resize(image, (ww, hh))
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
image = cv2.resize(image, (171, 128))
if is_flip == 1 and doFlip:
image = cv2.flip(image, 1)
if doTrans:
hh, ww = image.shape[0:2]
M = np.float32([[1, 0, tran_x], [0, 1, tran_y]])
image = cv2.warpAffine(image, M, (ww, hh))
if doNoisy:
noisy(image, 'gauss')
batch[i][j][:][:][:] = image[crop_x:crop_x + 112,
crop_y:crop_y + 112, :]
labels[i] = label
else:
for j in range(16):
img = imgs[symbol + j]
# image = cv2.imread(img_path + path + '/' + img)
image = cv2.imread(img)
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
image = cv2.resize(image, (171, 128))
batch[i][j][:][:][:] = image[8:120, 30:142, :]
labels[i] = label
return batch, labels
def siamese_process_batch(lines, img_path, train=True):
num = len(lines)
batch = np.zeros((num, 16, 112, 112, 3), dtype='float32')
btcof = np.zeros((num, 16, 112, 112, 3), dtype='float32')
labels = np.zeros(num, dtype='int')
for i in range(num):
path = lines[i].split(' ')[0]
label = lines[i].split(' ')[-1]
symbol = lines[i].split(' ')[1]
label = label.strip('\n')
label = int(label)
symbol = int(symbol) - 1
if not os.path.isfile(path + '/{:06d}'.format(symbol + 16) + '.jpg'):
continue
file_path = path + '/{:06d}'.format(symbol)
flow_ext = ''
use_npy = True
if isfile(file_path + '.npy'):
flow_ext = '.npy'
elif isfile(file_path + '.npz'):
flow_ext = '.npz'
else:
use_npy = False
flow_ext = '_flow.jpg'
if train:
crop_x = random.randint(0, 10)
crop_y = random.randint(0, 10)
tran_x = random.randint(0, 10)
tran_y = random.randint(0, 10)
is_flip = random.randint(0, 1)
sc = random.uniform(0.2, 1)
for j in range(16):
img_file = path + '/{:06d}'.format(symbol + j) + '.jpg'
image = cv2.imread(img_file)
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
if use_npy:
im_of = np.load(file_path + flow_ext)
else:
im_of = cv2.imread(file_path + flow_ext)
if im_of.shape[2] == 2:
mag, ang = cv2.cartToPolar(im_of[..., 0], im_of[..., 1])
im_of = np.dstack((im_of, mag))
# im_of *= 16
im_of = cv2.normalize(im_of, None, -1, 1, cv2.NORM_MINMAX)
# im_of = cv2.resize(im_of, (img_w, img_h))
# Re-scale
if doScale:
hh, ww = image.shape[0:2]
hh = int(hh * sc)
ww = int(ww * sc)
image = cv2.resize(image, (ww, hh))
im_of = cv2.resize(im_of, (ww, hh))
# image = cv2.resize(image, (171, 128))
if is_flip == 1 and doFlip:
image = cv2.flip(image, 1)
im_of = cv2.flip(im_of, 1)
if doTrans:
hh, ww = image.shape[0:2]
M = np.float32([[1, 0, tran_x], [0, 1, tran_y]])
image = cv2.warpAffine(image, M, (ww, hh))
im_of = cv2.warpAffine(im_of, M, (ww, hh))
if doNoisy:
noisy(image, 'gauss')
noisy(im_of, 'gauss')
# cv2.imshow('opticalflow',im_of[crop_x:crop_x + 112, crop_y:crop_y + 112, :])
# cv2.waitKey(1)
image = cv2.resize(image[crop_x:, crop_y:, :],
(kernel_w, kernel_h))
im_of = cv2.resize(im_of[crop_x:, crop_y:, :],
(kernel_w, kernel_h))
batch[i][j][:][:][:] = image
btcof[i][j][:][:][:] = im_of
labels[i] = label
else:
for j in range(16):
file_path = path + '/{:06d}'.format(symbol + j)
image = cv2.imread(file_path + '.jpg')
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
# if not os.path.isfile(file_path + flow_ext):
# print(file_path + flow_ext)
if use_npy:
im_of = np.load(file_path + flow_ext)
else:
im_of = cv2.imread(file_path + flow_ext)
if im_of.shape[2] == 2:
mag, ang = cv2.cartToPolar(im_of[..., 0], im_of[..., 1])
im_of = np.dstack((im_of, mag))
im_of = cv2.normalize(im_of, None, -1, 1, cv2.NORM_MINMAX)
image = cv2.resize(image, (kernel_w, kernel_h))
im_of = cv2.resize(im_of, (kernel_w, kernel_h))
batch[i][j][:][:][:] = image
btcof[i][j][:][:][:] = im_of
labels[i] = label
return batch, btcof, labels
def process_batch_opt_flow(lines, img_path, train=True):
num = len(lines)
btcof = np.zeros((num, 16, 112, 112, 3), dtype='float32')
labels = np.zeros(num, dtype='int')
for i in range(num):
path = lines[i].split(' ')[0]
label = lines[i].split(' ')[-1]
symbol = lines[i].split(' ')[1]
label = label.strip('\n')
label = int(label)
symbol = int(symbol) - 1
# imgs = os.listdir(img_path+path)
imgs = glob.glob(img_path + path + '/*.jpg')
if imgs == []:
imgs = glob.glob(img_path + path + '/*.png')
imgs.sort(key=str.lower)
if (symbol + 16) >= len(imgs):
continue
if train:
crop_x = random.randint(0, 10)
crop_y = random.randint(0, 10)
tran_x = random.randint(0, 10)
tran_y = random.randint(0, 10)
is_flip = random.randint(0, 1)
init_f = random.randint(0, 8)
init_f = 0
if (symbol + 16 + init_f) >= len(imgs):
init_f = 0
sc = random.uniform(0.2, 1)
for j in range(16):
img_file = imgs[symbol + j + init_f]
im_of = np.load(img_file + '.npy')
im_of *= 16
im_of = cv2.normalize(im_of, None, 0, 255, cv2.NORM_MINMAX)
# Re-scale
if doScale:
hh, ww = im_of.shape[0:2]
hh = int(hh * sc)
ww = int(ww * sc)
im_of = cv2.resize(im_of, (ww, hh))
# image = cv2.resize(image, (171, 128))
if is_flip == 1 and doFlip:
im_of = cv2.flip(im_of, 1)
if doTrans:
hh, ww = im_of.shape[0:2]
M = np.float32([[1, 0, tran_x], [0, 1, tran_y]])
im_of = cv2.warpAffine(im_of, M, (ww, hh))
if doNoisy:
noisy(im_of, 'gauss')
btcof[i][j][:][:][:] = im_of[crop_x:crop_x + 112,
crop_y:crop_y + 112, :]
labels[i] = label
else:
for j in range(16):
img_file = imgs[symbol + j]
# image = cv2.imread(img_path + path + '/' + img_file)
im_of = np.load(img_file + '.npy')
btcof[i][j][:][:][:] = im_of[8:120, 30:142, :]
labels[i] = label
return btcof, labels
def siamese_process_batch(lines, img_path, train=True):
num = len(lines)
batch = np.zeros((num, 16, kernel_w, kernel_h, 3), dtype='float32')
btcof = np.zeros((num, 16, kernel_w, kernel_h, 3), dtype='float32')
labels = np.zeros(num, dtype='int')
for i in range(num):
path = lines[i].split(' ')[0]
label = lines[i].split(' ')[-1]
symbol = lines[i].split(' ')[1]
label = label.strip('\n')
label = int(label)
symbol = int(symbol) - 1
# imgs = os.listdir(img_path+path)
imgs = glob.glob(img_path + path + '/*.jpg')
if imgs == []:
imgs = glob.glob(img_path + path + '/*.png')
imgs.sort(key=str.lower)
if (symbol + 16) >= len(imgs):
continue
if train:
# crop_x = random.randint(0, 15)
# crop_y = random.randint(0, 58)
# is_flip = random.randint(0, 1)
crop_x = random.randint(0, 10)
crop_y = random.randint(0, 10)
tran_x = random.randint(0, 10)
tran_y = random.randint(0, 10)
is_flip = random.randint(0, 1)
sc = random.uniform(0.2, 1)
for j in range(16):
# crop_x = random.randint(0, 15) # add detection noise
# crop_y = random.randint(0, 58)
img = imgs[symbol + j]
# image = cv2.imread(img_path + path + '/' + img)
image = cv2.imread(img)
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
# image = cv2.resize(image, (img_w, img_h))
im_of = np.load(img + '.npy') if isfile(
img + '.npy') else np.load(img[:-4] + '.npz')['flow']
if im_of.shape[2] == 2:
mag, ang = cv2.cartToPolar(im_of[..., 0], im_of[..., 1])
im_of = np.dstack((im_of, mag))
# im_of *= 16
im_of = cv2.normalize(im_of, None, -1, 1, cv2.NORM_MINMAX)
# im_of = cv2.resize(im_of, (img_w, img_h))
image = cv2.resize(image, (kernel_w, kernel_h))
im_of = cv2.resize(im_of, (kernel_w, kernel_h))
batch[i][j][:][:][:] = image
btcof[i][j][:][:][:] = im_of
continue
# Re-scale
if doScale:
hh, ww = image.shape[0:2]
hh = int(hh * sc)
ww = int(ww * sc)
image = cv2.resize(image, (ww, hh))
im_of = cv2.resize(im_of, (ww, hh))
if is_flip == 1 and doFlip:
image = cv2.flip(image, 1)
im_of = cv2.flip(im_of, 1)
if doTrans:
hh, ww = image.shape[0:2]
M = np.float32([[1, 0, tran_x], [0, 1, tran_y]])
image = cv2.warpAffine(image, M, (ww, hh))
im_of = cv2.warpAffine(im_of, M, (ww, hh))
if doNoisy:
noisy(image, 'gauss')
noisy(im_of, 'gauss')
batch[i][j][:][:][:] = image[crop_x:crop_x + kernel_w,
crop_y:crop_y + kernel_h, :]
btcof[i][j][:][:][:] = im_of[crop_x:crop_x + kernel_w,
crop_y:crop_y + kernel_h, :]
labels[i] = label
else:
for j in range(16):
img = imgs[symbol + j]
# image = cv2.imread(img_path + path + '/' + img)
image = cv2.imread(img)
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
im_of = np.load(img + '.npy') if isfile(
img + '.npy') else np.load(img[:-4] + '.npz')['flow']
if im_of.shape[2] == 2:
mag, ang = cv2.cartToPolar(im_of[..., 0], im_of[..., 1])
im_of = np.dstack((im_of, mag))
im_of = cv2.normalize(im_of, None, -1, 1, cv2.NORM_MINMAX)
image = cv2.resize(image, (kernel_w, kernel_h))
im_of = cv2.resize(im_of, (kernel_w, kernel_h))
batch[i][j][:][:][:] = image
btcof[i][j][:][:][:] = im_of
continue
image = cv2.resize(image, (img_w, img_h))
im_of = cv2.resize(im_of, (img_w, img_h))
h_start = (img_h - kernel_h) // 2
h_end = h_start + kernel_h
w_start = (img_w - kernel_w) // 2
w_end = w_start + kernel_w
batch[i][j][:][:][:] = image[h_start:h_end, w_start:w_end, :]
btcof[i][j][:][:][:] = im_of[h_start:h_end, w_start:w_end, :]
labels[i] = label
return batch, btcof, labels