Segmentation Models on artificial moon imagery

Oct 20, 2019
Code Video

In this project, I trained 4 deep learning segmentation models on an artificial Lunar Dataset to see how they will perform on real images from Nasa. For this project, I trained and tested 4 different segmentation models:

  • UNet
  • LinkNet
  • PSPNet
  • FPN
All of them had very similar training procedure, you can therefore consult the notebook I used to train the FPN and extrapolate the main components of it to the others. The second notebook is where I tested my model on the test dataset and on real moon images. This dataset comes from kaggle. I worked with:
  • Around 7000 images for train set
  • Around 2000 images for validation set
  • Around 1000 images for the test set
  • Around 40 images from real moon pictures
I then tried my model on an Apollo video shot from a rover driven during the 1972's Apollo 15 mission. All these results are consultable in my presentation video.

You can also dive into my code in these multiple notebooks !


Deep Learning Python Computer Vision ipynb ML DS
Avatar
Thomas Di Martino
PhD Student in AI & Remote Sensing

My research interests include deep learning technologies, automatic feature extraction and computer vision, all of them applied to Remote Sensing problematics, more precisely to Synthetic Aperture Radar (SAR) acquisitions.

Related