An attempt (though not a successful one) to predict truth or lie in a video or audio using machine learning. It is still under progress, we are working on better feature encoding and performance metrics to train classifier.
There are two folders one for audio based (MFCC features) attempt and the other is video based (DLIB 68 point features) attempt. Both folders have python scripts files "custom_data_evaluator.py" and "custom_test_score.py" respectively. Those two scripts use the already trained classifier for video and audio features to predict label for test videos located in "test_data" folder. Execute these two files to generate results we got.
1 - The data set is too much noisy.
2 - Video based features are not normalized for face scales and face orientation in video frames.
3 - Variable length data.
4 - Use of various sensors to capture videos and audios.
5 - Feature encoding used is wrong. (means and standard deviation of all frames of a video/audio are taken as one feature vector of that video)