Researchers built an aerial-object-detection dataset inside a war game.
The team behind G-MAD, an open-source framework, uses Arma3 - the military simulation game - to generate synchronized visible and thermal (RGB-T) imagery from multiple aerial viewpoints. The engine's internal geometry data automatically produces bounding-box annotations, skipping the manual labeling that normally eats up dataset-building budgets. G-MAD also lets researchers control camera placement and scenario details precisely, something real drone flights rarely allow. Using the framework, the team released AMOD, a new large-scale benchmark for multi-view aerial RGB-T detection, along with the source code.
Aerial object detection - the kind of computer vision that spots vehicles, people, or structures from drone or aircraft footage - has long been bottlenecked by data. Capturing synchronized visible and thermal footage from multiple angles in the real world is expensive, and aligning the two camera types perfectly is fiddly work. Game engines have solved similar problems before, most famously when researchers mined GTA V for synthetic self-driving data a decade ago; G-MAD applies the same trick to thermal aerial imagery, a domain with far less existing data to draw from.
The catch is the one every synthetic dataset faces: a model trained on rendered pixels still has to prove it works on real footage, and simulated thermal signatures are only a rough stand-in for actual heat.