We are keen on stickers, so at the beginning we have been thinking in make an app that makes a 8bits sticker for the phone, but none of us know much about mobile development.
What it does
The app is simple: you take a photo with your phone or another device and select it. Face2Sprite will analize your picture using deep learning algorithms and try to find your facial traits. The AI then will select the sprite that resembles you most accurately. Errors may occur, but we find that this adds a very funny random factor.
How we built it
We used various libraries but, mainly, fastai is the base of our project. We took a database from Kaggle.com (celebrities faces), adapted the data and select the fields that were most important to us. With that data, we trained the AI (architecture resnet50) with more than 200k photos in various epochs. Before that, we started with a pre-trained model (imagenet) and customized the model with our dataset. The .gifs were created with a community based software (Sprite Creator 3) that provided us the necessary tools to build the different models. The sprites where finally animated with PhotoScape.
Challenges we ran into
- Inexperience (this was the first hackathon we participated in)
- Lack of software, as we had to slightly change the original idea.
- Insufficient hardware, as training an AI is very demanding and we depended on a remote GPU.
- Lack of time, because doing a project in just 36 hours is just not easy.
Accomplishments that we're proud of
- Being able to finish the project.
- The commitment that we had on the project.
- The ability to work as a team.
What we learned
- To adapt the datasets to our necessities.
- The time it takes to built a project.
- The differences between thinking an idea and make it work.
- Do some sacrifices.
What's next for Face2Sprite
Being able to import the sprite into your favourite games.