Generative Adversarial Networks (GAN) for training Machine Learning algorithms
Introduction:
Build a GAN to improve the Machine Learning algorithm to identify products in PIM based on camera input.
The challenge of Machine Learning is often to get a training set large enough. This project aims to use General Adversarial Networks (GAN) to learn from one domain and use the generated visuals and apply it to another, in order to get a large number of (fake) images to train the next product's recognition data set.
Here is a simple example:
The bottles of Pepsi on the left is the starting point. And the 256 pictures to the right is the goal - pictures from every thinkable angle with obstacles in front on some pictures etc.
Now, if we can analyze the correspondance between the small set (4 samples, domain A below) and the large set (256 samples) perhaps we can use the same learning to start from another limited sample and use the GAN to create 256 new (fake) images (in domain B).
E.g. take 2 pictures of a Mountain Dew bottle like here and use what we learned from domain A:
Want to learn more? Download Rune's Master Thesis here>>
Rune Eggers
Aarhus University
Master of Data Science, Aarhus BSS