def _py_func_vp(rgb_path, flow_path, label):
rgb_path = rgb_path.decode ()
flow_path = flow_path.decode()
rgb_cap = cv2.VideoCapture(rgb_path)
flow_cap = cv2.VideoCapture(flow_path)
flow_len = flow_cap.get(cv2.CAP_PROP_FRAME_COUNT)
rgb_len = rgb_cap.get(cv2.CAP_PROP_FRAME_COUNT)
index = np.random.randint(0,min(rgb_len,flow_len) - 10)
rgb_cap.set(cv2.CAP_PROP_POS_FRAMES,index)
rgb_file = None
while rgb_file is None:
_ , image = rgb_cap.read()
if image is not None:
image = cv2.cvtColor(image,cv2.COLOR_BGR2RGB)
image = np.float32(image)
image = cv2.resize(image,(340,256))
rgb_file = DataAugmentation.Multiscale_crop(image)
rgb_file = DataAugmentation.horizontal_flip(rgb_file)
rgb_file = cv2.resize(rgb_file,(224,224))
rgb_file = normalize(rgb_file)
break
flow_cap.set(cv2.CAP_PROP_FRAME_COUNT,index)
image_list = []
for i in range(10):
f , img = flow_cap.read()
if img is not None:
image_list.append(img[...,:2])
image = np.concatenate(image_list,axis=-1)
image = np.float32(image)
image = cv2.resize(image,(340,256))
flow_file = DataAugmentation.Multiscale_crop(image,is_flow=True)
flow_file = DataAugmentation.horizontal_flip(flow_file,is_flow=True)
flow_file = cv2.resize(flow_file,(224,224))
flow_file = (flow_file / 255) * 2 - 1
one_hot_label = np.zeros (101, dtype=np.float32)
one_hot_label[label] = 1
return rgb_file, flow_file, one_hot_label
def _py_func_rgb_vp(rgb_path, label):
rgb_path = rgb_path.decode ()
rgb_file = os.listdir (rgb_path)
rgb_file = sorted(rgb_file)
try:
random_file = random.choice(rgb_file)
rgb_file_path = os.path.join(rgb_path,random_file)
img = cv2.imread(rgb_file_path)
img = cv2.cvtColor(img,cv2.COLOR_BGR2RGB)
except:
raise ValueError('rgb_file_path',rgb_file_path,'cannot be read')
img = cv2.resize(img,(340,256))
img = DataAugmentation.Multiscale_crop(img)
img = DataAugmentation.horizontal_flip(img)
img = cv2.resize(img,(224,224))
img = np.float32(img)
img = normalize(img)
assert img.shape == (224,224,3)
if label is not None:
one_hot_label = np.zeros (101, dtype=np.float32)
one_hot_label[label] = 1
return img, one_hot_label
return img
def _py_func_rgb_vp(rgb_path, label):
rgb_path = rgb_path.decode()
rgb_cap = cv2.VideoCapture(rgb_path)
rgb_len = rgb_cap.get(cv2.CAP_PROP_FRAME_COUNT)
while 1:
index = np.random.randint(0, rgb_len)
rgb_cap.set(cv2.CAP_PROP_POS_FRAMES, index)
_, image = rgb_cap.read()
if image is not None:
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
image = np.float32(image)
image = cv2.resize(image, (340, 256))
rgb_file = DataAugmentation.Multiscale_crop(image)
rgb_file = DataAugmentation.horizontal_flip(rgb_file)
rgb_file = cv2.resize(rgb_file, (224, 224))
if _preprocess_name == 'pytorch':
rgb_file = normalize(rgb_file)
elif _preprocess_name == 'tf':
rgb_file = tf_preprocess(rgb_file)
break
rgb_cap.release()
if label is not None:
one_hot_label = np.zeros(101, dtype=np.float32)
one_hot_label[label] = 1
return rgb_file, one_hot_label
return rgb_file
def _py_func_flow_vp(flow_path, label):
flow_path = flow_path.decode ()
flow_cap = cv2.VideoCapture (flow_path)
flow_len = flow_cap.get (cv2.CAP_PROP_FRAME_COUNT)
flag = True
image_list = []
try:
index = np.random.randint (0, flow_len - _frame_counts)
except:
index = 0
flow_cap.set (cv2.CAP_PROP_POS_FRAMES, index)
for i in range(_frame_counts):
flag, image = flow_cap.read ()
if image is not None:
image = np.float32 (image)
image = cv2.resize (image, (340, 256))
image_list.append (image[...,:2])
if len (image_list) == _frame_counts or flag is False:
break
flow_cap.release ()
image = np.stack (image_list, axis=0)
assert 0 not in image.shape
if _frame_counts != image.shape[0]:
image = np.repeat(image,(_frame_counts//image.shape[0])+1,axis=0)[:_frame_counts,...]
assert (_frame_counts,256,340,2) == image.shape
# flow_file = DataAugmentation.randomCrop (image, 112, 112)
# rgb_file = DataAugmentation.Multiscale_crop(image,is_flow=True)
# flow_file = DataAugmentation.randomCrop(image,224,224)
flow_file = DataAugmentation.borders25(image)
flow_file = DataAugmentation.horizontal_flip (flow_file)
flow_file = np.float32([cv2.resize(x,(224,224)) for x in flow_file])
# image_list = []
# for i in range(rgb_file.shape[0]):
# image_list.append(cv2.resize(flow_file[i],(224,224)))
# flow_file = np.stack(image_list,axis=0)
flow_file = np.float32(flow_file)
flow_file = (flow_file / 255) * 2 - 1
if label is not None:
one_hot_label = np.zeros (101, dtype=np.float32)
one_hot_label[label] = 1
return flow_file, one_hot_label
return flow_file
def _py_func_flow_vp(flow_path, label, _index_list=None):
f_upath, f_vpath = flow_path
flow_u_path = f_upath.decode()
flow_v_path = f_vpath.decode()
flow_file = os.listdir(flow_u_path)
flow_file = sorted(flow_file)
flow_len = len(flow_file)
if _index_list is None:
index_list = ucf_dataset.pick_index(flow_len - _new_length,
_frame_counts,
is_flow=True)
else:
index_list = _index_list
flow_img = []
for j in index_list:
img_list = []
for i in range(j, j + _new_length):
img_u_path = os.path.join(flow_u_path, flow_file[i])
img_v_path = os.path.join(flow_v_path, flow_file[i])
img_u = cv2.imread(img_u_path, 0)
img_v = cv2.imread(img_v_path, 0)
if img_u is None or img_v is None:
continue
img = np.stack([img_u, img_v], axis=-1)
img_list.append(img)
img = np.concatenate(img_list, axis=-1)
img = cv2.resize(img, (340, 256))
img = DataAugmentation.Multiscale_crop(img, is_flow=True)
img = DataAugmentation.horizontal_flip(img)
img = cv2.resize(img, (224, 224))
img = np.float32(img)
img = (img / 255 - 0.5) / 0.226
flow_img.append(img)
np.random.shuffle(flow_img)
if label is not None:
one_hot_label = np.zeros(101, dtype=np.float32)
one_hot_label[label] = 1
return flow_img, one_hot_label
return flow_img
def _py_func_flow_vp(flow_path, label):
flow_path = flow_path.decode ()
# flow_file = _flow_vp (flow_path)
flow_cap = cv2.VideoCapture(flow_path)
flow_len = flow_cap.get(cv2.CAP_PROP_FRAME_COUNT)
num_segments = 3
segments_list = np.array_split(np.arange(0,flow_len),num_segments)
flag = True
image_list = []
while num_segments:
index = np.random.choice(segments_list[num_segments-1])
flow_cap.set(cv2.CAP_PROP_POS_FRAMES,index)
img_list = []
for i in range(10):
f , img = flow_cap.read()
if img is not None:
img_list.append(img[...,:2])
if len(img_list) != 10:
continue
image = np.concatenate(img_list,axis=-1)
image = cv2.cvtColor(image,cv2.COLOR_BGR2RGB)
image = np.float32(image)
image = cv2.resize(image,(340,256))
image = DataAugmentation.Multiscale_crop(image,is_flow=True)
image = DataAugmentation.horizontal_flip(image)
image = cv2.resize(image,(224,224))
image = (image / 255) * 2 - 1
image_list.append(image)
num_segments -= 1
flow_file = image_list
flow_cap.release()
if label is not None:
one_hot_label = np.zeros (101, dtype=np.float32)
one_hot_label[label] = 1
return flow_file, one_hot_label
return flow_file
def _py_func_rgb_vp(rgb_path, label, _index_list=None):
rgb_path = rgb_path.decode()
rgb_cap = cv2.VideoCapture(rgb_path)
rgb_len = rgb_cap.get(cv2.CAP_PROP_FRAME_COUNT)
image_list = []
if _index_list is None:
index_list = ucf_dataset.pick_index(rgb_len,
_frame_counts,
is_flow=True)
else:
index_list = _index_list
for index in index_list:
rgb_cap.set(cv2.CAP_PROP_POS_FRAMES, index)
flag, image = rgb_cap.read()
if image is not None:
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
image = np.float32(image)
image = cv2.resize(image, (340, 256))
image = DataAugmentation.Multiscale_crop(image)
image = cv2.resize(image, (224, 224))
image = DataAugmentation.horizontal_flip(image)
image = normalize(image)
image_list.append(image)
if len(image_list) != _frame_counts:
image_list = np.concatenate([image_list] * _frame_counts,
axis=0)[:_frame_counts]
rgb_cap.release()
np.random.shuffle(image_list)
rgb_file = np.stack(image_list, axis=0)
if label is not None:
one_hot_label = np.zeros(101, dtype=np.float32)
one_hot_label[label] = 1
return rgb_file, one_hot_label
return rgb_file
def _py_func_rgb_vp(rgb_path, label):
rgb_path = rgb_path.decode()
rgb_cap = cv2.VideoCapture(rgb_path)
rgb_len = rgb_cap.get(cv2.CAP_PROP_FRAME_COUNT)
image_list = []
index = np.random.randint(0, rgb_len - _frame_counts)
index_list = np.arange(index, index + _frame_counts)
for index in index_list:
rgb_cap.set(cv2.CAP_PROP_POS_FRAMES, index)
flag, image = rgb_cap.read()
if image is not None:
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
image = cv2.resize(image, (340, 256))
image_list.append(image)
if len(image_list) != _frame_counts:
image_list = np.concatenate([image_list] * _frame_counts,
axis=0)[:_frame_counts]
assert image_list != []
rgb_cap.release()
image = np.stack(image_list, axis=0)
rgb_file = DataAugmentation.randomCrop(image, 224, 224)
rgb_file = DataAugmentation.horizontal_flip(rgb_file)
rgb_file = np.float32(rgb_file)
rgb_file = normalize(rgb_file)
if label is not None:
one_hot_label = np.zeros(101, dtype=np.float32)
one_hot_label[label] = 1
return rgb_file, one_hot_label
return rgb_file
def _py_func_rgb_vp(rgb_path, label):
rgb_path = rgb_path.decode ()
num_segments = 3
# rgb_file = _rgb_vp (rgb_path)
rgb_cap = cv2.VideoCapture(rgb_path)
rgb_len = rgb_cap.get(cv2.CAP_PROP_FRAME_COUNT)
segments_list = np.array_split(np.arange(0,rgb_len),num_segments)
image_list = []
while num_segments:
index = np.random.choice(segments_list[num_segments-1])
rgb_cap.set(cv2.CAP_PROP_POS_FRAMES,index)
flag , image = rgb_cap.read()
if image is None:
continue
image = cv2.cvtColor(image,cv2.COLOR_BGR2RGB)
image = np.float32(image)
image = cv2.resize(image,(340,256))
image = DataAugmentation.Multiscale_crop(image)
image = DataAugmentation.horizontal_flip(image)
image = cv2.resize(image,(224,224))
image = normalize(image)
image_list.append(image)
num_segments -= 1
assert len(image_list) != num_segments
rgb_cap.release()
rgb_file = image_list
if label is not None:
one_hot_label = np.zeros (101, dtype=np.float32)
one_hot_label[label] = 1
return rgb_file, one_hot_label
return rgb_file
def _py_func_flow_vp(flow_path, label):
flow_path = flow_path.decode ()
flow_cap = cv2.VideoCapture(flow_path)
flow_len = flow_cap.get(cv2.CAP_PROP_FRAME_COUNT)
flag = True
image_list = []
while flag:
try:
index = np.random.randint(0,flow_len-10)
except:
raise ValueError('flow len is too short with %d frame' % flow_len,flow_path)
flow_cap.set(cv2.CAP_PROP_POS_FRAMES,index)
for i in range(10):
f , img = flow_cap.read()
if img is not None:
image_list.append(img[...,:2])
if len(image_list) == 10:
flag = False
image = np.concatenate(image_list,axis=-1)
flow_cap.release()
image = np.float32(image)
image = cv2.resize(image,(340,256))
flow_file = DataAugmentation.Multiscale_crop(image,is_flow=True)
flow_file = DataAugmentation.horizontal_flip(flow_file)
flow_file = cv2.resize(flow_file,(224,224))
flow_file = (flow_file / 255 - 0.5) / 0.226
assert 0 not in flow_file.shape
if label is not None:
one_hot_label = np.zeros (101, dtype=np.float32)
one_hot_label[label] = 1
return flow_file, one_hot_label
return flow_file
def _py_func_flow_vp(flow_path, label):
f_upath, f_vpath = flow_path
flow_u_path = f_upath.decode()
flow_v_path = f_vpath.decode()
flow_file = os.listdir(flow_u_path)
flow_file = sorted(flow_file)
index = np.random.randint(0, len(flow_file) - _new_length)
img_list = []
for i in range(index, index + _new_length):
img_u_path = os.path.join(flow_u_path, flow_file[i])
img_v_path = os.path.join(flow_v_path, flow_file[i])
img_u = cv2.imread(img_u_path, 0)
img_v = cv2.imread(img_v_path, 0)
img = np.stack([img_u, img_v], axis=-1)
img_list.append(img)
img = np.concatenate(img_list, axis=-1)
img = cv2.resize(img, (340, 256))
img = DataAugmentation.Multiscale_crop(img, is_flow=True)
img = DataAugmentation.horizontal_flip(img)
img = cv2.resize(img, (224, 224))
img = np.float32(img)
if _preprocess_name == 'pytorch':
img = (img / 255 - 0.5) / 0.226
elif _preprocess_name == 'tf':
img = tf_preprocess(img)
if label is not None:
one_hot_label = np.zeros(101, dtype=np.float32)
one_hot_label[label] = 1
return img, one_hot_label
return img
def _py_func_flow_vp(flow_path, label):
f_upath, f_vpath = flow_path
flow_u_path = f_upath.decode()
flow_v_path = f_vpath.decode()
flow_file = os.listdir(flow_u_path)
flow_file = sorted(flow_file)
index = np.random.randint(0, len(flow_file) - _frame_counts)
img_list = []
for j in range(index, index + _frame_counts):
img_u_path = os.path.join(flow_u_path, flow_file[j])
img_v_path = os.path.join(flow_v_path, flow_file[j])
img_u = cv2.imread(img_u_path, 0)
img_v = cv2.imread(img_v_path, 0)
if img_u is None or img_v is None:
continue
img = np.stack([img_u, img_v], axis=-1)
img = cv2.resize(img, (340, 256))
img_list.append(img)
img = np.stack(img_list, axis=0)
img = DataAugmentation.randomCrop(img, 224, 224)
img = DataAugmentation.horizontal_flip(img)
img = np.float32(img)
img = (img / 255 - 0.5) / 0.226
if label is not None:
one_hot_label = np.zeros(101, dtype=np.float32)
one_hot_label[label] = 1
return img, one_hot_label
return img
def _py_func_rgb_vp(rgb_path, label):
rgb_path = rgb_path.decode ()
# rgb_file = _rgb_vp (rgb_path)
rgb_cap = cv2.VideoCapture (rgb_path)
rgb_len = rgb_cap.get (cv2.CAP_PROP_FRAME_COUNT)
image_list = []
# while 1:
try:
index = np.random.randint (0, rgb_len-_frame_counts)
except:
index = 0
rgb_cap.set (cv2.CAP_PROP_POS_FRAMES, index)
while 1:
flag, image = rgb_cap.read ()
if image is not None:
# image = np.float32 (image)
image = cv2.cvtColor(image,cv2.COLOR_BGR2RGB)
image = cv2.resize (image, (340, 256))
image_list.append(image)
if len(image_list) == _frame_counts or flag is False:
if len(image_list) == 0:
rgb_cap.set (cv2.CAP_PROP_POS_FRAMES, 0)
continue
break
rgb_cap.release ()
image = np.stack(image_list,axis=0)
# some bug in rgb video,maybe you should use ffmpeg for image ,but read video is faster than read image due to IO spare
assert 0 not in image.shape
if _frame_counts != image.shape[0]:
image = np.repeat(image,(_frame_counts//image.shape[0])+1,axis=0)[:_frame_counts,...]
assert (_frame_counts,256,340,3) == image.shape
if _crop == 'random':
rgb_file = DataAugmentation.randomCrop(image,224,224)
# print('random crop is used')
else:
rgb_file = DataAugmentation.Multiscale_crop(image)
# print('mutltisacle crop is used')
rgb_file = np.float32([cv2.resize(x,(224,224)) for x in rgb_file])
assert 0 not in rgb_file.shape
rgb_file = DataAugmentation.horizontal_flip (rgb_file)
rgb_file = np.float32(rgb_file)
rgb_file = normalize (rgb_file)
# rgb_file = (rgb_file / 255) * 2 - 1
assert rgb_file is not None
if label is not None:
one_hot_label = np.zeros (101, dtype=np.float32)
one_hot_label[label] = 1
return rgb_file, one_hot_label
return rgb_file