Problem

Does a registration network trained on MNIST digits carry over to unseen QuickDraw shapes?

What I did

Trained and evaluated Fourier-Net and Fourier-Net-Diff on MNIST and tested transfer to unseen shapes using Dice overlap.

Role

Application, training and evaluation of an existing model (from the original Fourier-Net authors), run on a university HPC cluster. I also wrote the introduction, background and conclusion.

Credits

Team project (CS 6501 Machine Learning in Image Analysis).