Wackipedia

Example generated article.
Welcome page of the website.
Another example article.

Inspiration

This was our first hackathon, so we wanted to work on something fun and weird. We were inspired by Planet Earth documentaries, which are narrated by David Attenborough. We wanted to create an app that would allow us to input any image or video and have it be narrated by David Attenborough as if it were part of the Planet Earth series.

What it does

This is the first step toward our goal. In its current form, Wackipedia takes in an image from the user. It then determines what the main object in the image is and generates a fake Wikipedia article about it. If an object was not detected with high confidence, we would use the metadata of the image to extract the location that it was taken and generate an article about the city instead.

How we built it

We used a pre-trained model (trained on ImageNet) that would detect objects in the image. Then, we scraped articles about everyday or common objects from Wikipedia and fed them as biases to gpt-2, which allows us to generate short paragraphs of fake information in a style similar to Wikipedia articles. Finally, we host it on a python server that interacts with the HTML/JS.

Challenges we ran into

We tried hosting the app on Google Cloud, but it used over the allowed 2GB. Also, our initial idea was to extract the location data from the photo if no objects were detected, and write a wikipedia page about the city the photo was taken at instead, but we couldn't find a way to load multiple models simultaneously on tensorflow. Another source of trouble was generating the articles, since that takes a minute or so on our computer, which made testing what worked and what didn't take a while.

Accomplishments that we're proud of

We mostly accomplished what we set out to do, which is pretty cool!

What we learned

This was our first Hackathon, and we learned a lot! None of us had ever made web apps before, so this was something completely new. We learned how to hack together a project with many aspects together quickly, and how to adjust our goals as we go along.

What's next for Wackipedia

We would like it to be able to narrate videos in the future, in a similar way. Using IBM's Watson, we could use text-to-speech to read out what's happening in David Attenborough's voice.