tool · draft
PyTorch-Wildlife and MegaDetector
Open-source wildlife detection models from Microsoft's AI for Good Lab, including MegaDetector for camera-trap images, and PyTorch-Wildlife, the framework that runs those models.
What is reusable: open sourcepartly open dataAPI unknownunknown participation MIT (code); weights vary by variant (MIT, Apache-2.0, AGPL-3.0 or CC BY-NC-SA 4.0)
PyTorch-Wildlife is a framework for building, sharing and running conservation models. The lab's Biodiversity hub groups it with:1
- MegaDetector, for camera-trap images1,2
- MegaDetector-Acoustic, for sound1
- MegaDetector-Overhead, for drone and aerial imagery1,3
- a model zoo of species classifiers1
A camera trap is a motion-triggered camera left out in the field. MegaDetector draws boxes around animals, people and vehicles in its images, each with a confidence score.2 It does not identify species.2 Researchers use it mainly to discard empty frames before review.2 Dan Morris developed versions 1 to 5 at Microsoft.2 Version 6 uses the YOLOv9, YOLOv10 and RT-DETR architectures.2
MegaDetector-Overhead, formerly called OWL (Overhead Wildlife Locator), applies the same approach to drone survey imagery, where targets are small and altitude and perspective vary.3 Its models include a checkpoint for caribou.1,3 No-code front ends such as AddaxAI, SPARROW Studio and a Hugging Face demo can also load the models.2
Everything runs locally.2 pip install PytorchWildlife installs a Python API and a command line that write JSON detections per image.1,2 The code is MIT-licensed. Each variant's weights carry their own licence, and the overhead model and data release is CC BY-NC-SA 4.0.2,3
- form
- open-source framework and models
- group
- Wildlife and field monitoring
- homepage
- https://microsoft.github.io/Biodiversity/
- licence
- MIT for the code; model weights vary by variant (MIT, Apache-2.0, AGPL-3.0 or CC BY-NC-SA 4.0)
- cost
- free
Tags: wildlife, camera trap, detection, conservation, open source, computer vision