DroneMegaDetector

The detector is general-purpose, for aerial wildlife monitoring.^1,2 It is the Medium size of RF-DETR, an object-detection architecture, with three classes: person, animal and vehicle.¹ In practice "animal" means mammal.¹ It was built by Chaim Chai Elchik, Steven Longmore, Dan Morris, Ben Weinstein, Josh Veitch-Michaelis, Camee Knops, Freya Richards and Serge Wich.^1,2

It outputs bounding boxes, not points, so it can support counting and tracking as well as detection.² Most of the training data comes from drone flights at roughly 50 to 120 m above ground, which the authors call the principal operating range.¹ Higher up, animals appear smaller and detection falls off.¹ In the preprint it matched the OWL model on a cattle dataset and performed almost identically on caribou.² It transferred poorly to birds.² On high-altitude SAVMAP imagery it degraded, as did every other method.²

The code is MIT-licensed Python.¹ There are scripts for single images, for folders (writing MegaDetector-format JSON), for sampled video frames and for tracking with ByteTrack, a multi-object tracker.¹ Large images are split into tiles for inference.¹ A hosted copy is on Hugging Face as ConservationDrones/DroneMegaDetector.¹ The weights and code are released; the paper is a bioRxiv preprint, with version 2 posted on 14 October 2026.^1,2