Empowering users with Image colorization

Inspiration

Wayfair has many product images, and the cost of producing each image is high, since it usually involves multiple manual stages. These stages can be:

1) Top level product-image spec from stakeholder/supplier (1 to many days);

2) Initial stylist sketch (20 min to 2 hours);

3) Artist tweaks, lights, chooses material, and renders the image (20 min to 3 hours);

4) QA with top artist/stylist (goes into queue/backlog, ~10 min per review);

5) Final approval (~10 min).

On the occasion that we want to slightly modify the end result, for example in the case of a supplier manufacturing the same furniture item in a different color, we have to repeat the above pipeline, which costs $$$.

In this challenge day, we address the post processing stage. Specifically, the use case of a color change to part of or the entire scene. We propose a post-processing colorization tool that, which, using recent advances in deep learning, allows the interactive modification and colorization of the scene, which save time and speeds up the curation process.

In addition to benefiting creative personal, our tool empower teams, and stakeholders of the Wayfair product image stack.

What it does

Given a greyscale image, a user can interactively colorize the image using a color palette learned from a dataset. The result look realistic and preserves original image features.

How I built it

With pytorch and python. We built on top of previous research, open-source code available here: https://richzhang.github.io/ideepcolor/

Challenges I ran into

Compilation, gathering interior scene images, debugging machine learning code

Accomplishments that I'm proud of

See challenges, and video: https://www.youtube.com/watch?v=X2FX0mkBtBE

What I learned

There are many possible improvements and next steps. One future step is, for example, enabling more color choices for users. One way to achieve this is by training on Wayfair interior scene images.

What's next for Empowering users with Image colorization

Hopefully, productionizing this project as a tool for Wayfair creative professional, and other interested parties.

Built With

• python
• pytorch