Mass Polling System

Tired of sending "How do you feel about this lecture/meeting" google forms that nobody uses? Turn boring documents into an interactive webcam experience!

Comment

Inspiration

This was inspired by feedback / satisfactory forms that many users would skip due to the time needed to complete them. While these can be more detailed, they require the user to deeply remember and reflect on what happened in the lecture/meeting instead of just showing their emotions in the now.

What it does

There are multiple steps:

  1. The website takes the number of people going to be in the webcam, then takes the photo of everyone.
  2. A classification neural network finds each human in the image and splits the image up accordingly, with their thumbs up / down.
  3. Another classification model will indicate whether the human is gesturing thumbs up or thumbs down.
  4. These will get added to a tally, which when complete, gets sent back to the end user.

How we built it

We built the frontend using SvelteKit, which allowed us to easily make the website interactive with little code.

The website fetched to an Express server which acted as the middleman between the model and the frontend. The server would run the python file and return the results back.

Challenges we ran into

  • The initial thumb classification model was inaccurate, so a lot of time was taken into training the best one we could. We eventually used Google's Teachable Machine to create the model, with around 8000 images of training data.

  • Unfortunately we learnt that because of the training data resolution, the photos taken had to be in 360×360, which limited how many people can actually fit. We believe that if we were able to find a more diverse array of training data, we could have a model which works with any appropriate resolution.

  • We also realized that the processing time for each image was very slow (~5 seconds), but this was because Tensorflow on Windows only uses CPU. With a good GPU, we should be able to get it at least under a second.

Accomplishments that we're proud of

  • The thumb classification model is pretty good!
  • We found the UI to be modern and simple to use, and also supporting dark mode.

Built With

Share this project:

Updates