Tensorflow implementation for 'Temporal 3D ConvNets'(t3d)

'T3D' model can be applied to action recognition.Paper Url :https://arxiv.org/abs/1711.08200
Here is the tensorflow implementation with tf.slim to make code leaner. Different from the authors' original implementation.I convert traditional 3D convolution(eg. tf.nn.conv3d ) to P3D('pseudo-3d') convolution which can greatly reduce the number of parameters.
(P3D details:http://openaccess.thecvf.com/content_ICCV_2017/papers/Qiu_Learning_Spatio-Temporal_Representation_ICCV_2017_paper.pdf )

Suppose you are about to use UCF dataset.Firstly converting videos to images is necessary. To do this,you could run codes like follows:(Suppose UCF-101 dataset is in the same directory as the code-files.)

• ./process_video2image.sh UCF101
• And next step,you should get the 'train.list' and 'test.list' which you would afterwards fetch from for training data and testing data individually:(number ‘5’ indicates that one-fifth of all data is testing data.)
• ./process_gettxt.sh UCF101 5

Note that:Due to the fact that Relative Path of the video clips exist in 'train.list' and 'test.list', So you must make sure that 'DataGenerator.py' and UCF-101 are in the same directory! or modify the codes by yourself.

After getting your own data and setting preference parameters in 'settings.py',You can run python train.py --txt='./train.list'(input parameter 'txt' has default value:'./train.list',in this way you can just run python train.py) to train model. You can also train and test model in 'tf-p3d-train_eval.ipynb' with jupyter notebook.

• Changing the properties for data augmentation in 'DataAugmenter.py'

• DO NOT use tf.contrib.layers.batch_norm() or slim.batch_norm() which may lead to wrong answers when testing.Using tf.layers.batch_normalization(training=...) instead.