On the main page README you mention:
"Is there any technical improvement used in this round than previous ones?
To train models for Australia we only had a few thousand building labels, which made it hard to rely only on supervised training. Typically we’ve used hundreds of thousands or best case tens of millions of building labels for training. In order to create a good and robust model for Australia we took advantage of self-supervised training and unsupervised domain adaptation techniques to leverage our training data from other countries and domains. We believe this is a good proof of concept to scale to building extraction to the whole world."
Do you have any papers or more details that you can point to on the self-supervised and unsupervised domain adaption techniques you used?
On the main page README you mention:
"Is there any technical improvement used in this round than previous ones?
To train models for Australia we only had a few thousand building labels, which made it hard to rely only on supervised training. Typically we’ve used hundreds of thousands or best case tens of millions of building labels for training. In order to create a good and robust model for Australia we took advantage of self-supervised training and unsupervised domain adaptation techniques to leverage our training data from other countries and domains. We believe this is a good proof of concept to scale to building extraction to the whole world."
Do you have any papers or more details that you can point to on the self-supervised and unsupervised domain adaption techniques you used?