Downloads · 30 days
0
dinov3s-buildings
hotosm/dinov3s-buildings
dinov3s-buildings is a machine learning model from hotosm. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as cc-by-4.0.
Building footprint segmentation for very-high-resolution aerial and satellite imagery. A frozen DINOv3 ViT-S/16 backbone with a trainable UperNet decoder, used in HOTOSM fAIr to extract building polygons from RGB imag…
Access
Public
Updated Jun 29, 2026
Repo size
894 MB
Likes
4
Public
Repo makeup
Click a slice to open those files.
.ckpt674 MB · 75%
At a glance
- License
- cc-by-4.0
- Access
- Public
- Created
- Jun 29, 2026
- Updated
- Jun 29, 2026
- SHA
- f7254545
Base models
Datasets
Spaces
Model card
From the Hugging Face model README
kshitijrajsharma/dinov3s-buildings
Building footprint segmentation for very-high-resolution aerial and satellite imagery. A frozen DINOv3 ViT-S/16 backbone with a trainable UperNet decoder, used in HOTOSM fAIr to extract building polygons from RGB imagery.
Inputs and outputs
- Input: RGB image tile.
- Output: per-pixel building probability, thresholded and vectorised to building polygons.
Inference
- Full-resolution continuous sliding window: 256 px window, 192 px stride.
- Default probability threshold:
0.4371. - Polygons are vectorised from the probability mask and simplified.
Files
model.onnx: self-contained inference graph (RGB tile -> building probability).model.ckpt: Lightning checkpoint for evaluation or further training.
Backbone
DINOv3 ViT-S/16, frozen. The UperNet decoder and segmentation head are the trained parameters.
Training data
Trained on hotosm/vhr-building-segmentation
(CC-BY-4.0 / ODbL).
License
CC-BY-4.0, following the training data. The DINOv3 backbone weights are subject to the separate DINOv3 license from their original release.
- License
- cc-by-4.0
- Created
- Jun 29, 2026
- Updated
- Jun 29, 2026