Generative Models as a Data Augmentation for Classification

Use GAN steerability as an data augmentation technique.

In this project, we use GAN steerability as an data augmentation technique. The inspiration is coming from GAN steerability, and GenRep these two papers. We investigate image transformation by exploring walks in the latent space of GAN.

With a generator G and magnitude $\alpha$, we try to learn latent vectors $W_{steer}$ , which achieves the same effects of transformation $T$ in the image space. As a consequence, we can manipulate the latent space to do transformation. We implement 3 transformation, which are rotation, zoom and shift and color transformation.

For experiment, we compared the results of using only real data, real data + augmented data in image space, real data + augmented data in the latent space (GAN steer), and only generated data in the latent space. And we conclude that GAN steerability is a better data augmentation technique compare to transformation done in the data space

More information of our implementation:
Code:
Video:
Slides: