Author: Jiaqing Lin E-mail: jiaqing930@gmail.com Code for paper”Tow-Stream Convolutional Networks for Action Recognition in Videos”

In this folder includes video data and source code. Folders: 1. “videos” folder has 4 classes video, basketball, biking, diving, and volleyball, each class includes 2 video file as .avi format (size is about 150 - 300 KB). 2. “rgb_images” and “flow_images” are prepared to store converted images. 3. The training log file is stored in “spatial_result” and “temporal_result” folders. Because training and test data are too small, so result is not good enough. Files: 1. “labels.txt” is stored each class label. 2. “result.png” is shown snapshot when running code in terminal. Code: 1. “read_data.py” is used to convert video data to image data for training models. 2. “spatial_model.py” includes spatial stream model, training, and evaluating. 3. “temporal_model.py” includes temporal stream model, training, and evaluating.

Each layer parameters is set by paper “Two-Stream Convolutional Networks for
Action Recognition in Videos” and other related papers, but parameters of training
is set by myself.

Library version: Python version 3.6.1 Chainer version 1.23.0 OpenCV version 3.2.0 Numpy version 1.12.1 Video dataset is download from youtube video dataset

Running code step: 1. : > python3 read_data.py (convert video data to each image data) 2. : > python3 spatial_model.py (training and evaluating) 3. : > python3 temporal_model.py (training and evaluating)