/humanana

A Tensorflow image classification model trained using Keras on around 5,000 pictures to accurately predict whether an object identifies more similar to a human or banana with 89% confidence after processing.

Primary LanguagePureBasicMIT LicenseMIT

Humanana

Ever wondered if you look more like a human or a banana? Here's your chance to go bananas and find out.

Inspiration

Humans. Bananas. Predator. Prey. So different, yet so alike; truly natural wonders of the modern world.

Humans worldwide eat over 100 million bananas per year. In fact, we're willing to bet that you ate a banana or two at HackTheNorth alone. Such a trivial and often overlooked fruit has demonstrated its greatness, and we believe that there is more to discover.

For example, did you know that humans and bananas share 50-60% of our DNA? Most people tend to agree that humans and bananas look nothing alike, but what if there's more to it than the eye can see? What if we look more like bananas than we think?

To answer all of these questions, we used nearly 5,000 images to train an image classification model that can find the truth in the most banana-like human (or human-like banana). Welcome to Humanana.

What it does

Given a picture from your camera or photo library, Humanana (read: "human nah nah") tells you whether you look more like a banana or a human and how confident it is. Come by and check out our live demo!

How we built it

(too many banana pictures) + (too many human pictures) = a dataset of nearly 5000 images, which we used to train our image classification model.

Challenges we ran into

  • bananas were very confidently identified as people
  • people were very confidently identified as bananas
  • grayscale images do not have colour, which interfered with our model
  • iOS apps do not use Python

Accomplishments that we're proud of

  • Humanana outputs its results rather quickly (< 2 seconds)
  • the user interface is almost exactly what we'd imagined

What we learned

  • whether we looked more like humans or bananas
  • bananas constitute at least 25% of our team
  • how to train our own machine learning models
  • how long it takes to download 30 000 images
  • an image's background colour is more important to the model than the object's colour

What's next for Humanana

  • humanatee
  • humonkey
  • humanchego cheese
  • humailbox
  • huMan Francisco
  • humackathon ... and many more!