Questions for pretrained models

[YJ-He]

Hi,

Thank you for sharing the outstanding work to the public, your contribution is worth to appreciate too much. When I try to use your shared pretrained weights to perform inference, I found a very interesting thing. The pretrained weights from "inria_dataset_polygonized.unet_resnet101_pretrained.leaderboard | 2020-06-02 07:57:31" could produce more regularized building footprints (see figure below), but other three pretrained models such as “mapping_dataset.unet_resnet101_pretrained.train_val | 2020-09-07 11:28:51” cannot do the same thing. So I'm wondering that if you perform any regurlization operations when you train the model "inria_dataset_polygonized.unet_resnet101_pretrained.leaderboard | 2020-06-02 07:57:31" . Why the building boundaries look so regular?

Thanks