Two-stream and convLSTM

• Description :This is the sorce code associated with the paper 《Two-Stream Convolutional Network for Improving Activity Recognition Using Convolutional Long Short-Term Memory Networks》
• File Description：
  • CNN-GPUs：this is a floder contains implement of cnn used as Two-stram network, suporting to fine-tune on specific dataset with multi-GPU
  • CNN-Pred-Feat:use the networks defined in the CNN-GPUs to extract both RGB feature description and optical feature description
  • Optical-flow:use opencv to compute optical information and map the optical information into optical video
  • RNN-convLSTM: the implement of convLSTM and the tarining and testing process of convLSTM on the feature maps from two-stram CNN
• Dependencies:
  • torch
  • cudnn8.0 and CUDA
  • extracunn
  • RNN library
  • ffmpeg
  • Torch Video Decoder Library