Inspiration
We wanted to challenge ourselves with something new. Something we felt fixed an issue we've observed in our every day lives. With the rising costs of living, many cities face a homeless crisis that they're struggling to manage. We sought to provide a way to enable the average person to more easily contribute to providing shelter and essential services to the un-homed in a flexible and direct way.
What it does
Our software banks the unbanked, using facial recognition to attach a credit account to the undocumented. DonAIte enables anybody to send funds to a registered account, using their face as an account reference. This credit can be used to pay for shelter and vital services, ensuring your donated money is spent in a manner that you'd be proud.
How we built it
We split into two group. The first building the application and framework, whereas the second built and trained a deep convolutional neural network to calculate unique facial embeddings.
Challenges we ran into
The AI team struggled wrestling large datasets and preprocessing them for the correct training tuples. whereas the systems team, with their mostly back-end experience, felt challenged by building front-ends for the first time.
Accomplishments that we're proud of
The AI team are proud of their model, and preprocessing scripts. We compared our model to another implementation we found, only to discover our model produced bigger differences in embeddings between different individuals, supporting the argument that out model produces more unique embeddings for each individual.
The systems team and the ai team are both so proud about how much work we got done within the fleeting time. We worked through the night on our work, consistently persevering in the face of technical adversity.
What we learned
The AI team learnt just how challenging a facial recognition model is to train. We learnt the importance of minimising biases in training data, and the need for multithreaded data processing libraries when dealing with large datasets.
What's next for DonAIte
As we only managed to train our recognition system on only one subset of our selected dataset, we plan to fine tune and improve its reliability with the rest. Also, with a better fine tuned model, we could register new donatees by detecting that they do not match any other users registered in the database.
Log in or sign up for Devpost to join the conversation.