Inspiration

People lie everyday from simple deception games like Among Us to lying in court. Therefore, we were thinking about creative ways we could approach this problem from a computer science standpoint.

What it does

It's not completely working yet since our preprocessing took a bit longer than expected, but what it intended to do was create and train a LSTM network based off of defendant data from our dataset. From there, we called the google api to process facial information, and pass that information as tensors into the network.

How I built it

We used Flask to host our server with a JS front end. Additionally, we used the python google api and PyTorch in order to create the tensors and model respectively.

Challenges I ran into

In terms of training time, our data took far longer than we expected since we had to sort through the data to make sure to get the quality parts out. In addition, we underestimated the server complexity in terms of debugging. Though some of us had work with the server end before, the unexpected debugging time took far too long and we ran up against the deadline.

Accomplishments that I'm proud of

Though we did not complete our project, we did gained a lot more experience with front end, back end connection with pytorch and google api, but most importantly, we're really proud of the work we did to clean up the dataset to make everything usable and be trained accurately.

What I learned

Throughout the project, we learned how to use several frontend, backend, and ML tools. For web development, we learned how to create a basic website with HTML and Javascript. On the machine learning side, we learned how to use the Google vision API and apply LSTM models to video data.

What's next for Lie Detection

We would increase accuracy of our product by finding higher datasets with less noise (i.e. interruptions or obstruction of facial features). We would also like to expand our model to incorporate audio features as we believe they could aid the lie detection process. Finally, we like to deploy onto accessible website where people can test the model against themselves with sample videos.

Built With

Share this project:

Updates