我在Keras有这个Conv3D模型:
model = Sequential(
Conv3D(32, (3,3,3), activation='relu', input_shape=self.input_shape),
MaxPooling3D(pool_size=(1, 2, 2), strides=(1, 2, 2)),
Conv3D(64, (3,3,3), activation='relu'),
MaxPooling3D(pool_size=(1, 2, 2), strides=(1, 2, 2)),
Conv3D(128, (3,3,3), activation='relu'),
Conv3D(128, (3,3,3), activation='relu'),
MaxPooling3D(pool_size=(1, 2, 2), strides=(1, 2, 2)),
Conv3D(256, (2,2,2), activation='relu'),
Conv3D(256, (2,2,2), activation='relu'),
MaxPooling3D(pool_size=(1, 2, 2), strides=(1, 2, 2)),
Flatten(),
Dense(1024)),
Dropout(0.5),
Dense(1024),
Dropout(0.5)),
Dense(self.nb_classes, activation='softmax')
)
此模型基于本文https://arxiv.org/pdf/1412.0767.pdf
使用此Conv3D预处理视频数据的最佳方法是什么?
我写了这个函数来从UCF-101的每个视频中提取帧:
def frame_writer(pathIn, pathOut, class_name):
"""
This function will read videos and write frames in a new dataset
args:
pathIn -> base dataset of videos
pathOut -> destination folder for the frames ('data/path')
"""
#creating output path if it not exists
try:
if not os.path.exists(pathOut + '/' + class_name):
os.makedirs(pathOut + '/' + class_name)
else:
pass
except:
print('Invalid path!')
#getting the list containing all files from the directory
pathIn_files = glob.glob(pathIn + '\\' + class_name + '\\' + '*.avi')
video_limit = len(pathIn_files)
#iterating over all files
for i, j in zip(pathIn_files, range(len(pathIn_files))):
#getting the names from file paths
base_name = os.path.basename(pathIn_files[j])
file_name = base_name[0:-4] #taking only the file name (without extension)
#getting the frames
vidcap = cv2.VideoCapture(i)
success,image = vidcap.read()
count = 0
success = True
while success:
success,image = vidcap.read()
print ('Read a new frame: ', success)
cv2.imwrite(pathOut + '\\' + class_name + "\\%s_frame%d.jpg" % (file_name, count), image)
count += 1
print('Done!')
现在我的框架数据集如下:
文件夹:数据
-SUBFOLDER:火车
- SUBFOLDER:class1
--- frame1_video1_class1.jpg
--- frame2_video1_class1.jpg
--- frame3_video1_class1.jpg
...
--- frameN_videoN_class1.jpg
- SUBFOLDER:class2
--- frame1_video1_class2.jpg
--- frame2_vide1_class2.jpg
--- frame3_video1_class2.jpg
...
--- frameN_videoN_class2.jpg
-SUBFOLDER:测试
- SUBFOLDER:class1
--- frame1_video1_class1.jpg
--- frame2_video1_class1.jpg
--- frame3_video1_class1.jpg
...
--- frameN_videoN_class1.jpg
- 子文件夹:class2
--- frame1_video1_class2.jpg
--- frame2_video1_class2.jpg
--- frame3_video1_class2.jpg
...
--- frameN_videoN_class2.jpg
所以我拥有与其类相对应的文件夹内所有视频的所有帧。
我必须使用keras函数中的ImageDataGenerator将它传递给我的Conv3D模型吗?
那么,在这种情况下,每次从每个班级传递每个视频的每一帧?
或者我必须以另一种方式做到这一点?
我只需要使用此模型预测视频!
感谢您的支持!
答案 0 :(得分:1)
一种方法是将所有帧放入一个大张量,相应地标记它们,并将其用作Keras模型的输入。张量中的帧数将是您的批量大小。