What it does?

  • Triplet loss to encourages dissimilar alignments be distant from similar alignments
  • Siamese network and triplet loss can be combined with different input types. For example; as in the previous slide, the rotation change is more understandable in sliced MIP data. Thus, feature extraction phase can be more successful.

How we built it?

  • Step by step:
  • Data Analysis
  • Data Load
  • Design
  • Training
  • Validation
  • Repeat – 3

Challenges we ran into

  • Generalization on the unseen radar sensors
  • Understanding data in short notice

Accomplishments that we're proud of

  • Ideas that we bring into table
  • Teamwork
  • Free swag that we collected

What we learned

  • It is our first experience with the Microwave Imaging dataset
  • We get to know new companies and people

What's next for

  • Improving the Augmentation methods
  • Working on different approaches for generalize the network better to get accurate alignment results on the unseen radar sensors
  • Attention can be used for extracting relations better. Attention is all you need

Built With

Share this project:

Updates