Project catalogue¶
The Microsoft AI for Good Lab keeps each biodiversity project in its own repository. This page helps you pick a starting point, shows how the projects connect, and links the model zoo. The home page has the short list of repositories.
Which project should I use?¶
| You want to | Start here | Also useful |
|---|---|---|
| Try the tools on your own images, videos or audio | SPARROW Playground: notebooks and scripts, no training | Quickstart notebook |
| Find animals, people and vehicles in camera-trap images | Sparrow Engine with a MegaDetector model from the zoo (pip install sparrow-engine) |
MegaDetector documentation: model variants, output format and confidence thresholds |
| Name the species in camera-trap detections | Sparrow Engine with a camera-trap classifier from the zoo | Detection and classification notebook |
| Train or fine-tune a species classifier on your own data | MegaDetector-Classifier | |
| Detect or identify species in audio recordings | Sparrow Engine with an acoustic model from the zoo | MegaDetector-Acoustic (the lab's bioacoustic models); bioacoustics notebook |
| Detect wildlife in drone or aerial images | MegaDetector-Overhead (OWL models) | Overhead models in the zoo |
| Run models from the command line, over HTTP, or from C, C++, Go, Rust or C# | Sparrow Engine: spe CLI, Docker HTTP server, C library |
User manual |
| Receive data from SPARROW devices, 4G e-mail cameras, the Reconyx cloud or FTPS cameras | Sparrow Comm (pip install sparrow-comm) |
Field-data notebooks |
| Deploy a solar-powered AI device at a remote site | SPARROW (AI edge device with 4G or satellite uplink) | SPARROWmini (small nodes that relay over a radio mesh to a SPARROW gateway, or upload directly over Wi-Fi or 4G); SPARROW documentation |
| Add your model to the Sparrow Model Zoo | SPARROW Model Zoo Uploader: convert, check outputs against the original, submit | |
| Use or read the old PyTorch-Wildlife code | The archive/pytorch-wildlife branch of SPARROW Playground (unmaintained) |
From PyTorch-Wildlife notebook: maps old classes and calls to Sparrow Engine |
How the projects fit together¶
- Capture. A SPARROW device, a SPARROWmini node or a commercial camera trap records images or audio in the field.
- Receive. Sparrow Comm collects that data over the device API, e-mail, the Reconyx cloud or FTPS.
- Analyse. Sparrow Engine runs models from the Sparrow Model Zoo on it: MegaDetector to find animals, then a classifier for species, or acoustic models for sound.
- Explore. SPARROW Playground shows each step in a notebook you can copy into your own code.
Model development happens in the model repositories: MegaDetector, MegaDetector-Classifier, MegaDetector-Acoustic and MegaDetector-Overhead. SPARROW Model Zoo Uploader converts a trained model, checks its outputs against the original, and submits it to the zoo that Sparrow Engine runs.
Model zoo¶
The Sparrow Model Zoo holds the lab's models and third-party models for camera traps, acoustics, overhead and marine imagery, and general use. Each model keeps its own licence, so check it before use, especially for commercial work.
- Catalogue: the Sparrow Engine model zoo catalogue lists every model with its source, licence and citation. Sections: camera-trap detectors, camera-trap classifiers, acoustic detectors, acoustic classifiers, overhead detectors, marine imagery detectors, licensing at a glance.
- Model files: the Sparrow Model Zoo on Zenodo. This DOI always opens the latest version.
- MegaDetector accuracy and model sizes: the MegaDetector model zoo page.
- Earlier weights: the AI4G-Opossum classifier from PyTorch-Wildlife is not in the Sparrow Model
Zoo. Its original weights (
OpossumClassification_v0.0.0.ckpt) stay on Zenodo record 10023414.
The model zoo pages that used to live on this site (model_zoo/megadetector/,
model_zoo/classifiers/, model_zoo/other_detectors/, model_zoo/bioacoustics/) now redirect to
the matching sections of the catalogue.
Get involved¶
Ask questions on Discord. Bug reports and feature requests go to the repository of the project they concern; see Contribute. The teams already using these tools are on the collaborators page.