C3D-Action-Recognition

Train the C3D network with my own data set. Video or gif can be supported as a training file. Video streams or image frames can be used as input for detection.

Environment

• opencv-3.2
• keras-2.0.8
• tensorflow-1.3

Train your own data

• Place the data in the datasets/ucf101. The label for training (train_file.txt), label for testing (test_file.txt), are placed in the /ucfTrainTestlist. Record category in classInd.txt.

• Run video2img.py and save it in the datasets/ucfimgs. Currently gif and video formats are supported.

• Run make_label_txt.py and generate train_list.txt and test_list.txt.

• Modify model.py, such as lines 7~9.

• Modify train_c3d.py, such as lines 158~164. Run train_c3d.py.

Demo

• Modify config.txt
  classInd_path : the file path of the record category
  lr : learning rate (default 0.005)
  momentum : Momentum (default 0.9)
  weights_path : path to the trained model
  

  Image mode:
  image_read : the path to the image read
  image_write : the path written by the image
  

  video_image: select input mode
  video: the mode of video input
  image: the mode of image input
  
  

• Modify video_demo.py
  If you want to apply it to the video stream, you must first modify the dictionary(video_stream).
  video_stream[key].[0] : This is an array that records 16 video frames as input to the model.
  video_stream[key].[1] : the path to the data.
  video_stream[key].[2] : the path of the model.
  
  

• Run video_demo.py
  Identification of abnormal-event and normal-event:
  
  Identification of smoking: