feichtenhofer / twostreamfusion

Code release for "Convolutional Two-Stream Network Fusion for Video Action Recognition", CVPR 2016.

================================================================================

Convolutional Two-Stream Network Fusion for Video Action Recognition

This repository contains the code for our CVPR 2016 paper:

Christoph Feichtenhofer, Axel Pinz, Andrew Zisserman
"Convolutional Two-Stream Network Fusion for Video Action Recognition"
in Proc. CVPR 2016

If you find the code useful for your research, please cite our paper:

    @inproceedings{feichtenhofer2016convolutional,
      title={Convolutional Two-Stream Network Fusion for Video Action Recognition},
      author={Feichtenhofer, Christoph and Pinz, Axel and Zisserman, Andrew},
      booktitle={Conference on Computer Vision and Pattern Recognition (CVPR)},
      year={2016}
    }

Requirements

The code was tested on Ubuntu 14.04 and Windows 10 using MATLAB R2015b and NVIDIA Titan X or Z GPUs.

If you have questions regarding the implementation please contact:

Christoph Feichtenhofer <feichtenhofer AT tugraz.at>

================================================================================

Setup

1. Download the code git clone --recursive https://github.com/feichtenhofer/twostreamfusion

2. Compile the code by running compile.m.

   • This will also compile a modified (and older) version of the MatConvNet toolbox. In case of any issues, please follow the installation instructions on the MatConvNet homepage.

3. Edit the file cnn_setup_environment.m to adjust the models and data paths.

4. Download pretrained model files and the datasets, linked below and unpack them into your models/data directory.

• Optionally you can pretrain your own twostream models by running
  1. cnn_ucf101_spatial(); to train the appearance network stream.
  2. cnn_ucf101_temporal(); to train the optical flow network stream.

5. Run cnn_ucf101_fusion(); this will use the downloaded models and demonstrate training of our final architecture on UCF101/HMDB51.
   • In case you would like to train on the CPU, clear the variable opts.train.gpus
   • In case you encounter memory issues on your GPU, consider decreasing the cudnnWorkspaceLimit (512MB is default)

Pretrained models

• Download our baseline networks trained on UCF101 here:
  • VGG-16
  • ResNet-50
  • ResNet-152

Data

Pre-computed optical flow images and resized rgb frames for the UCF101 and HMDB51 datasets

• UCF101 RGB: part1 part2 part3

• UCF101 Flow: part1 part2 part3

• HMDB51 RGB: part1

• HMDB51 Flow: part1

Use it on your own dataset

• Our Optical flow extraction tool provides OpenCV wrappers for optical flow extraction on a GPU.