Sketchers

Inspiration

Sketching artists may spend a large amount of time gathering evidence on suspects' appearances from witnesses. They also spend a considerable amount of time perfecting their representations of suspects so that they may look more realistic. We would like to automate much of that process so that artists can start from a descriptive body sketch and a realistic image of a face rather than sketch everything from scratch.

What it does

Input: answers from a witness describing the suspect

1) (Current) Displays visual body description as questions are answered

2) (Future) Generates realistic face photos from sketches.

How I built it

• MongoDB for the database
• Node js for server
• React for the client
• Google Cloud for MongoDB storage
• Photoshop for designing base sketches of bodies and clothes.
• (In progress) Tensorflow + Pytorch + OpenCV for realistic face photo generation from sketches.

Challenges I ran into

• Many of us were not familiar with react so we took more time to learn about it.
• We attempted to use a pre-trained ML model but were met with multiple hurdles. We made a simple PCA based model but its results did not satisfy us (in fact they were scary 0_0). We attempted to use pre-trained models from recent research projects like CA-GAN, but we had trouble with transferring large files to the GPU environment we had access to, which impacted our ability to run the models.

Accomplishments that I'm proud of

• We are proud of aiming for something bold that we liked and believed in.

What I learned

• React
• Mongoose
• and many many ML models for the sketch to realistic photo generation.

What's next for Sketchers

• The next phase is for us to complete the ML part with a pre-trained model.

Built With

• machine-learning
• node.js
• react