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.