Trix with Pix

Project Checkin #1: https://docs.google.com/document/d/1vLKHusfXASHFV191XZINUW7ml7wrxXutN2qjZ_aQ9W0/edit?usp=sharing

Project Checkin #2: https://docs.google.com/document/d/1Ny0ImPwHDrZBRv8l7tRwqbLgJ7KQKQE_QxI8GEWke_A/edit?usp=sharing

Poster: https://drive.google.com/file/d/1km1TytzQvmzO6WEJ4L1DMc7PdBanJWk4/view?usp=sharing

Oral Presentation: https://drive.google.com/file/d/1fdUQJ39q1G0lRFVWxwQ3b2dCTz5wVoY-/view?usp=sharing

Final writeup/reflection: https://docs.google.com/document/d/111ZuHRBD4tzYvYMRkx8N4lNIwfTt0lnEVyWM-xmDVmg/edit?usp=sharing

Code: https://github.com/christopherbravo/Pix2Pix-Colorizing-Landscapes

Inspiration

We used Insola et al. 2018 for inspiration of how to use cGANs to solve image-to-image translation problems.

What it does

Our model converts blueprints of buildings to realistic images of building facades.

How we built it

We coded a cGAN with the same architecture as the original Insola et al. paper in Tensorflow.

Challenges we ran into

• We messed up the order of data in preprocessing
• We spent significant time revising our loss function
• We spent time to understand the model architecture, especially how the U-Net and skip connections work.

Accomplishments that we're proud of

We were able to get model to run :) Even thought it wasn't perfect our model did create faint building images that were somewhat recognizable. We also have a deep understanding of the loss function and the underlying probabilities.

What we learned

We learned how skip connections work to preserve low level information in the generator.

What's next for Trix with Pix

We plan to continue working on our model to improve performance on the original pix2pix datasets and experiment with it on new datasets.

Built With

• excel
• powerpoint
• prezi