When we trained our baseline HoVer-Net model, we used ImageNet ResNet50 pretrained weights. However, this was missing when we released the training code, which explains why participants could not quite reproduce the results. The training code (conic branch of HoVer-Net repository) is now exactly in line with our own implementation (including data split and augmentations). You can find the code here.

You can see the training code fix in this PR.