ORBITaL-Net: A labeled training library for large-scale building feature extraction.

There is a clear trend over time (and thus expansion of the global training dataset in quantity and variety of samples) towards models that are more generalizable according to both the increase in overall F1 score and the narrowing of the ratio between precision and recall scores, a general indicator that a model has learned a more “stable” prediction function.

also in 19,460 other papers