Environment:

• Ubuntu 16.04
• Python: 2.7
• openCV: 3.0
• Cuda: 9.1

Packages: Please refer to:

Contain all the packages and environment configurations that need to be set before runnning any script I highly recommend you run the script mannually line by line to avoid any error You must run the script in the right order to avoid dependency problem

This part contains all files related with data processing procedure.

Executables:

A end-to-end pipeline that process the raw videos and generate video features. The end file will be videoFeatures.tar.gz Argument: N/A

Automatically download the raw video and unzip them Argument: N/A

Automatically download the raw audio and unzip them

Remove video that has no corresponding audio Argument: N/A

Match onset and duration of video with audio Argument: N/A

Calculate the duration of the audio Argument: a .wav file

Extract number of frames of an audio Argument: a .wav file

Extract all frames from a video into jpg files Argument: N/A

Use pretrained face detector to extract ROI Argument:

• shape-predictor: a pretrained facial landmark predicator (will be downloaded by the ./dataProcessing.sh if not exist)
• image-Path: path to all images. Notice it is not the path of images, but the folder that contains all images
• outputPath: where the extracted mouths are going to Usage: python mouthExtract.py -p -i -o

Use pretrained googLeNet to extract features Argument:

• text file containing directories to images
• text file where the features will be written to Usage: python featureExtract.py -i -o

Conduct PCA on the features. Pick out 63 features and normalize them. Argument:

• Input vec file: a file contains features of every video
• Output vec file: a file contains the after-processed features of every video

Write each frame's feature to the corresponding file Argument:

• entire.vec: a file containing all after-pca features
• output.list: a file containing the output path corresponding to the entire.vec

Duplicate each feature by four times to match the frames of video and audio. Also take care of extra or missing frames. Argument: N/A