OWL inference notebook and sample images

OWL inference notebook

The notebook demonstrates pretrained OWL-D on four aerial image patches, with optional tiled inference on two full-resolution images. It shows environment/model details, progress, predicted counts, heatmaps, point overlays, and CSV/image exports.

Notebook: https://github.com/microsoft/MegaDetector-Overhead/blob/owl-notebook-v1/notebooks/owl_inference_demo.ipynb

GitHub previews the notebook. Clone the matching repository revision to run it locally in Jupyter or VS Code:

git clone --branch owl-notebook-v1 https://github.com/microsoft/MegaDetector-Overhead.git
cd MegaDetector-Overhead
uv sync --locked --no-default-groups --group gpu --extra notebook
.venv/bin/jupyter lab notebooks/owl_inference_demo.ipynb

Install Git and uv first. Default OWL-D requires a compatible NVIDIA CUDA GPU and driver. For CPU, install with uv sync --locked --extra notebook and explicitly select MODEL='owl-c', DEVICE='cpu'. No author-specific Conda environment is required. The notebook backend runs in the checkout's own locked .venv.

Use Restart Kernel and Run All Cells. Enable RUN_FULL_RESOLUTION=True to include the two larger images. The local Jupyter/VS Code route is supported; Colab remains experimental and has not been verified in a hosted runtime.

Sample data

Download OWL_SAMPLE_DATA.zip or let the notebook fetch it from the manifest URL. The archive contains exactly six unannotated images and is distributed as a release asset, not in Git history.

• Size: 18,269,117 bytes
• SHA-256: 0627f30d55a79f83a029cc574ba6af54669b561197b1a5fa123d4ddbd8fead8b
• SHA256SUMS and SAMPLE_ATTRIBUTION.json accompany the archive.

Source-specific terms:

• SheepCounter, DJI-named patches: Public Domain, as confirmed by the contributor. Source: https://universe.roboflow.com/riisprivate/sheepcounter . No specific CC0 instrument is inferred.
• HerdNet-associated data, two full-resolution JPGs: CC BY-NC-SA 4.0, as confirmed by the contributor. Original dataset and creator attribution: https://doi.org/10.58119/ULG/MIRUU5 . License: https://creativecommons.org/licenses/by-nc-sa/4.0/ .
• Caribou data, CAH-named patches: CC BY-NC-SA 4.0, verified from the primary release. Source: https://zenodo.org/records/20802844 . The release requests citation of Chacon et al. (2026), "Overhead Wildlife Locator (OWL): Benchmarking Weakly Supervised Learning for Aerial Wildlife Surveys", https://arxiv.org/abs/2606.13911 .

Retain original attribution, license links, and applicable change notices. The exact supplied archive bytes are unchanged; the manifest documents individual image hashes and source associations. The repository's MIT software license does not license the sample images.

Models and interpretation

OWL-D and OWL-C checkpoints remain on Zenodo record 20802844, whose release declares CC BY-NC-SA 4.0. The included DINOv3 backbone has additional terms documented in the repository. Downloaded model files are checked against published sizes and SHA-256 values. No separate Meta backbone download is required for OWL-D inference.

Counts are retained localization peaks, not validated population estimates. FIDT heatmap sums are not animal counts and peak scores are not calibrated probabilities. The samples have no ground truth, so the demo does not report accuracy metrics or species identification.