July 8, 2020

Image generation

We thought it would be interesting to generate new car models based on existing car images, or generate a new bicycle model (for example for Gazelle) based on existing bicycle images.
We started this project by gathering car images and trained a GAN (Generative Adversarial Network) which should get better at generating car images after training for a while.

We used 17.000 car images and trained the GAN for approximately 7 hours. After training and actually testing how well the model was trained, it was sometimes hard to see what the GAN actually generated. Although, some pictures are actually looking quite good (and some of them look very bad, obviously).

We didn’t really have a business case in mind where we could apply image generation, so it just remained an experiment. Also, if we want to make this a success we should have a bigger dataset (with better resolutions) to train with, or there should be a better pre-trained GAN available that we can train further with our own images.

  • Wilte Falkena

    Senior Data Scientist
  • Stephan Schrijver

    DataOps Engineer