This is the code for testing part of our CVPR code and further.

To run the code, you need to install ffmpeg and opencv 2.4.11.

We have attached a modified version of ffmpeg 2.7.2 on the code package.

To install ffmpeg and opencv. We follow the guide on http://www.ozbotz.org/opencv-installation/ . But you need to install ffmpeg of our version and opencv 2.4.11 instead.

Note This is a third-party website. We have no control of the content on this website. Please use the content carefully.

• For extracting Motion Vector image, you can run the extract_mvs_sample.sh. The code is designed for the videos with resolution of 320 * 240 (width * height). If you need to extract motion vector from different resolution, you can modified line 55-56 of draw_flow.cpp from MV-code-release folder to the desired resolution.

  Note The time reported in the paper is calculated without I/O, which means without the time of writing dump file and images to disk.

• The MV-code-release folder is created using eclipse.

The model of UCF-101 is released (both spatial part and motion vector part).

1. Motion vector model and Spatial Model for UCF-101 split1, split2, and split3 can be found on: https://drive.google.com/drive/folders/1RXXAauI_inR1KvO8ygQmeNBBKRbj3xDJ?usp=sharing

2. The images of motion vector extracted from UCF-101 can be found on: https://drive.google.com/open?id=1MKy7uDzXCF0Q-ZIkABpaPYIjHbVHDBwx

If you find any questions in this motion vector CNN code, please contact me: 1023zhangbowen@tongji.edu.cn zbwglory@gmail.com.

Best regards,

Bowen

Please cite our paper in your publications if it helps your research:

@inproceedings{ZhangWWQW16,
  author    = {Bowen Zhang and Limin Wang and Zhe Wang and Yu Qiao and Hanli Wang},
  title     = {Real-time Action Recognition with Enhanced Motion Vector CNNs},
  booktitle = {CVPR},
  year      = {2016},
  pages     = {2718-2726}
}

This code is distributed under MIT LICENSE