Skip to content
#

isprs-vaihingen

Here are 2 public repositories matching this topic...

Weakly semi-supervised semantic segmentation of ISPRS Vaihingen aerial images: 3 pixel-labelled images + crop-level tags. Tag loss, tag-constrained self-training and UniMatch-style consistency lift mIoU from 0.714 to 0.742 (pixel-label upper bound: 0.765). PyTorch.

  • Updated Oct 13, 2026
  • Python

Add this topic to your repo

To associate your repository with the isprs-vaihingen topic, visit your repo's landing page and select "manage topics."

Learn more