Inspiration
The inspiration for our project was our general interest in computer vision and our shared experiences with being mislabelled as POC. In addition, we thought it would be interesting to teach a machine to behave somewhat like a human. As all humans like to put each other in categories whether they mean to or not. In addition, Ethnicity is a big part of all our identities and experiences. By knowing someone's identity we may sometimes feel more connected through similar cultural backgrounds.
What it does
Using the webcam camera of a computer, RacialProfiler uses computer vision to lock in and identify individual faces and facial features. Using OpenCV, and other libraries to pre-train datasets, RacialProfiler is capable of displaying an individual's potential ethnicity based on the image it is given/captured.
How we built it
The project was split into 4 parts:
- Setting up computer vision with the ability to identify face and facial features
- Rounding up raw data as images to form datasets including a wide range of ethnicities to feed into the AI
- "Teaching" the AI to identify characteristics of ethnic backgrounds
- Linking the computer vision done in part one to the ai who had learned from the dataset in part 2, which would finally map faces captured in the video to specific ethnicities. Combining all 4 parts together, we were able to produce a bare minimum viable product
Challenges we ran into
The entire team was new to AI and Machine learning concepts and a lot of self-learning was required to create the project. Additionally, teaching the AI to identify features of certain backgrounds was a challenge and took up the bulk of our time overcoming the many errors that arose. We found that it is hard to categorize the ethnic background of individuals that hail from certain areas like North America, as they are highly diverse and would conflict with the datasets of other ethnic groups. Thus, we did not account for racial ambiguity, as well as gender identities that fall on the LGBTQIA+ spectrum.
Accomplishments that we're proud of
We are proud that we stepped out of our comfort zones to participate in a hackathon that challenged us in many different ways but led to a unique project. Plus, we learned many new aspects about computer vision and teaching machines about datasets, finally outputting a functional project.
What we learned
We learned how to use libraries like pickle, pillow, OpenCV, and os to teach AI in order to detect facial features. In addition, we learnt that we need a larger and more comprehensive dataset with greater diversity within ethnic groups. Furthermore, it is hard to avoid bias, when creating a racial profiling algorithm, as many ethnic groups share similar features, take west and east Africans as an example.
What's next for RacialProfiler
A lot more work is needed to produce more accurate ethnic background readings which would require more data fed to the AI. Additionally teaching the AI smaller more marginalized ethnic groups (ex: Central Asia) and overall more data to account for racially ambiguous individuals will significantly improve the machine. I believe as a next step we should try to implement deep learning for AI to accurately depict the knowledge it learns through the datasets it is given.
Log in or sign up for Devpost to join the conversation.