MegaDetector-Overhead (OWL)

MegaDetector was built for camera-trap photos, and this version adapts it to aerial surveys, with their small targets, varying altitude and straight-down views.¹ It is part of the PyTorch-Wildlife ecosystem.¹ The repository ships training, evaluation and inference code for three OWL (Overhead Wildlife Locator) models:

• OWL-C, a fast convolutional baseline.
• OWL-T, which adds a Swin transformer branch for cluttered backgrounds.
• OWL-D, built on a DINOv3 vision transformer.^1,2

The OWL paper (Chacón and others, June 2026) frames the models as weakly supervised density estimation.² They are trained from a single point marked on each animal instead of a drawn box, and estimate how many animals fall in each part of an image.² The paper benchmarks them on five public aerial datasets plus a caribou census transfer test.² A separate preprint by Elchik and others compares a general-purpose box detector, DroneMegaDetector, against OWL and finds performance varies by dataset and altitude.³

The code is MIT-licensed and runs locally in Python.¹ The OWL-D checkpoint is about 3.5 GB and needs a CUDA GPU, whereas OWL-C runs on CPU.¹ Model weights and the caribou datasets are on Zenodo under CC BY-NC-SA 4.0, and the README records a full local run on 8 September 2026.¹