CNN video analysis

Action recognition using ensemble of deep convolutional neural networks

     Deep convolutional neural networks (DCNN) 用在video上，

     通过视频帧间的时间信息（如光流、光流的梯度信息和图像的梯度信息），来区分不同的动作。

     光流的maps作为DCNN的输入。

Large-scale video classification with convolutional neural networks     

      Extend the connectivity of CNN in time domain to take advantage of local spatio-temporal information. 

      What temporal connectivity pattern in a CNN architecture is best at using local motion information of the video?

      为了加速，把CNN架构氛围两个过程来处理：

        （1）a context stream that learns features on low-resolution frames

         （2）a high-resolution fovea stream that only operates on the middle portion of the frame.

     每一个视频看做a bag of short, fixed-sized clips....

          Extend  the connectivity of the network in time dimension to learn spatio-temporal features 

          可以总结为一下三种情况： 

        

3D CNN （3D convolutional neural networks for human action recognition）