Posters presented at LPS22:
- A simplified model for building damage detection (xView2): PDF
- A potential problem in self-supervision on satellite images: PDF
I'm a second year PhD student, working with Josephine Sullivan.
My broad area of research is self-supervised learning for satellite images. Currently, I'm investigating the effect of noisy positive pairs in domain-specific contrastive learning on the downstream task performance.
For classification, it's detrimental to teach a network that a dog is a cat if you want it to distinguish them later on. For segmentation, it seems like it's not detrimental to tell the network that a whole building and a destroyed building are the same. That's pretty unexpected.
Here's my poster I'm presenting on Tuesday at LPS22 about this: Link to poster.
Deep Learning in Data Science (DD2424), assistant | Course web
Deep Learning, Advanced Course (DD2412), assistant | Course web