AMOD* Arma3 Multi-view Object Detection

Aerial object detection (AOD) has advanced rapidly due to its significant potential in various industries. Recent advances in visual sensors with deep learning have led to remarkable improvements in AOD performance. Yet, creating a well-organized dataset is labor-intensive and heavily costly. Synthetic data is a promising alternative, but how to construct and utilize it in aerial image understanding remains poorly understood. In this work, we aim to uncover what makes synthetic data truly useful for aerial object detection. We introduce Arma3 Multi-view Object Detection in aerial imagery (AMOD), a new benchmark built with a game Arma3, featuring multi-view visible aerial imagery annotated with oriented bounding boxes. Using our AMOD dataset, we find that simultaneous multi-angular viewpoints force the model to learn geometrically consistent and generalizable representations.

Below is the Illustration of how to construct our dataset named AMOD. (a) Geometric definition of the observing angle θ, where the camera shifts laterally by Δy in Y-axis, producing an off-nadir view. (b) Drone camera sweeping setup simultaneously capturing the same scene from nadir to very off-nadir viewpoints. (c) Examples from AMOD under different observation angles, showing how target appearance and perspective distortion vary with θ.