Mask2Former-DINOv2 — Drone Semantic Segmentation (9-class)

A Mask2Former semantic segmentation model with a DINOv2 ViT-B/14 backbone, full fine-tuned (no LoRA) for 9-class land-cover segmentation of high-resolution drone / orthomosaic imagery.

• Best validation mIoU: 74.45% (checkpoint mask2former-dinov2-drone-seg-9cls.pth, iter 24 000)
• Input: RGB image tiles at 1024px
• Output: Per-pixel semantic class map (9 classes; head emits 9 + "no object")
• Backbone: facebookresearch/dinov2 ViT-B/14 (vit_base_patch14_dinov2)
• Architecture: facebookresearch/Mask2Former + DINOv2 backbone
• Training code: https://github.com/vivekyaadav/mask2former-dinov2-drone-segmentation

Trained on a private, custom-annotated dataset. The dataset is not released.

Class taxonomy (0-indexed)

The predicted class-index map maps directly to this table; 255 = ignore.

Intended use

Semantic segmentation of urban/rural land-cover — building types, water bodies, and small infrastructure (tanks, wells, solar panels, vehicles) — from top-down drone or orthomosaic imagery. Research / GIS prototyping only; not validated for safety-critical use.

How to use

This is a Detectron2 + Mask2Former model with a custom DINOv2 backbone and patched matcher/criterion. You need the training repo (for the config and the patched fork), not just stock Mask2Former.

Recommended — end-to-end inference. For inference on a raw drone orthomosaic (TIFF → shapefile output), use scripts/ortho_pipeline.py from the training repo: github.com/vivekyaadav/mask2former-dinov2-drone-segmentation. It handles GSD resampling, tiling, inference, and vectorization automatically, and writes a styled GeoPackage + per-class shapefiles. This is the path that matches how the model is actually run in production.

First, set up the fork + patches (see the training repo's environment_setup.sh) and download the checkpoint:

python - <<'PY'
from huggingface_hub import hf_hub_download
ckpt = hf_hub_download(
    repo_id="vivekyaadav/mask2former-dinov2-drone-segmentation",
    filename="mask2former-dinov2-drone-seg-9cls.pth")
print(ckpt)
PY

Alternative — raw per-tile inference. If you only want the low-level Detectron2 predictor on individual 1024px tiles (no geospatial resampling/tiling/vectorization), use DefaultPredictor directly:

import cv2, torch
from detectron2.config import get_cfg
from detectron2.engine import DefaultPredictor
from detectron2.projects.deeplab import add_deeplab_config
from mask2former import add_maskformer2_config
from mask2former.config import add_dinov2_config   # from the patched fork

cfg = get_cfg()
add_deeplab_config(cfg)
add_maskformer2_config(cfg)
add_dinov2_config(cfg)
cfg.merge_from_file("configs/drone/drone_dinov2_semantic_v6.yaml")  # from the training repo
cfg.MODEL.WEIGHTS = ckpt
cfg.MODEL.DEVICE = "cuda" if torch.cuda.is_available() else "cpu"

predictor = DefaultPredictor(cfg)
image = cv2.imread("your_tile.png")          # BGR; 1024px recommended
sem = predictor(image)["sem_seg"].argmax(0).cpu().numpy()  # HxW class-index map (0..8)

Training details

Limitations

• Trained on a specific regional distribution; performance on other geographies, sensors, or resolutions is untested.
• Dataset size and class balance are not disclosed; treat performance as indicative.
• Heavy class imbalance (building_rcc ≈ 84% of pixels); rare classes such as overhead_tank remain comparatively weak (IoU ≈ 36).

Citation

@article{oquab2023dinov2,
  title={DINOv2: Learning Robust Visual Features without Supervision},
  author={Oquab, Maxime and others},
  journal={arXiv preprint arXiv:2304.07193},
  year={2023}
}

@inproceedings{cheng2022mask2former,
  title={Masked-attention Mask Transformer for Universal Image Segmentation},
  author={Cheng, Bowen and others},
  booktitle={CVPR},
  year={2022}
}

License

MIT. Underlying DINOv2 (Apache 2.0) and Mask2Former (MIT) licenses apply to those components.

Results

Best validation mIoU: 74.45% (iter 24 000)

Per-class validation IoU at the best checkpoint (iter 24 000):

Inference results