WayClear

Inspiration

3D model of a product provides us flexibility in creating images with various backgrounds in a short time with low cost (A 3D model takes $150 to generate, while an image takes $800 to be shot) & lesser manual labour. They are built using the supplier provided images that have a resolution of atleast 2000x2000. Obviously, not all images can be used to create 3D models due to poor quality. Thus, our motivation was to increase resolution of images to make them viable for 3D model building.

What it does

We Used a deep learning model (SR-CNN), that reads the low resolution model as input & converts them to a high resolution image, which can now be used to create 3D models.

How we built it

We decided to solve this problem for the sofa class. We started by collecting high resolution sofa images and downsampled them to make the dataset. The downsampled low resolution images became the input to the model and the output of the model was used to compare to the high resolution images. The model was trained to minimize the differences between the reconstructed images and the high resolution images.

Challenges we ran into

In the literature we referred to, most of the images were of 32x32 resolution and were being scaled up to 128x128 resolution. We tried to scale images up to 2000x2000 resolution which was very challenging. We also had very limited time & computational resources.

Accomplishments that we're proud of

We were proud that we were able to work on a business problem (Product imagery), completely unrelated to our day to day work, & provide a prototype with limited time & resources.

What we learned

We learnt a lot about how Deep Learning models (specially SR-CNN & SR-GAN) are applied to Super Resolution problems. We also learnt a lot about the image pipeline within Wayfair, including image sources, how images are created, processed & stored.

What's next for WayClear

Expand the model to train on multiple classes Experiment with more powerful models (e.g. ResNet), with different configurations (e.g. changing the loss function).