Jumping Into the Deep Learning Bandwagon

It turns out I completed my PhD in 2012 just before Deep Learning started to boom and be the Next Big Thing in Computer Vision. The various pieces were already there (neural networks, back propagation, large scale databases, GPGPU…), but still, I somehow managed to slip through it. Later at work, I’ve used standard Pattern Recognition approaches for realtime applications, and later the more complex and efficient Gradient Boosted Trees, but still, no luck at trying Deep Learning.

So, let’s dive in by setting up a pet project!

In this series of posts, I’ll log my path into having a working system that segments images. In order to get on-boarded faster, I’ve chosen to:

  • reuse an existing architecture: U-Net
  • implement it using PyTorch.

Then we’ll use that network to segment images into some classes to be defined (right now, I’m hesitating between the Dstl Kaggle challenge to create aerial maps or a deep version of Automatic Photo Pop-up) .

So, let’s dive in! First post to come in the series: Why U-Net?