Guangming Zhu

@JayNormanBaldwin Linux下写个shell脚本批量处理就是了

@emily19941102 The feature map which has the largest activation sum among the 256 channels is visualized. ConvLSTM每一步输出的都是二维的feature map，选取激活值之和最大的那个feature map，用matlab或其它程序可视化出来就行了。3DCNN也只能二维可视化某一时刻的特征图。

@emily19941102 可视化出来是二维图片，即使可视化成三维图片也只能看到立体的表面数据，你可以把3DCNN的每个时刻的特征图分别二维可视化，然后把可视化图片排列开来展示各个时刻的特征图就行了。

如果GPU的利用率会降为0，那表示CPU供数据速度不够，可以考虑把数据放到SSD上、多线程同时读数据、换用更快的视频解码库、事先把视频解码成图片等。

You can obtain the RGB/Depth/Flow images of IsoGD from the link: https://pan.baidu.com/s/12qC5sVhQAPizpsr0axVehw Password: fymm

@sumeetssaurav This work is done with the old version of Tensorlayer. You can refer to the new version using keras in TF-1.2: https://github.com/GuangmingZhu/ContinuousGR and https://github.com/GuangmingZhu/AttentioninConvLSTM.

Read README carefully. You need do some modification to tensorflow.contrib.keras.api.keras.layers.

@rex-yue-wu MIT license. The LICENSE file has been added.

写个程序对全部数据或部分数据进行统计计算得到

@JianweiDong (1) The keras.layers.GatedConvLSTM2D was modified from keras.layers.ConvLSTM2D by myself, it is just released in this repo. (2) I have no idea why the loss is nan. My previous researches...