Aerial Monocular 3D Object Detection

Drones equipped with cameras can significantly enhance human's ability to perceive the world because of their remarkable maneuverability in 3D space. Ironically, object detection for drones has always been conducted 2D image space, which fundamentally limits understand scenes. Furthermore, existing methods developed autonomous driving cannot be directly applied due lack deformation modeling, is essential distant aerial perspective sensitive distortion and small objects. To fill gap, this work proposes a dual-view system named DVDET achieve monocular both space physical address severe view issue, we propose novel trainable geo-deformable transformation module that properly warp information from drone's birds' eye (BEV). Compared cars, our includes learnable deformable network explicitly revising deviation. dataset challenge, new large-scale simulation AM3D-Sim, real-world AM3D-Real high-quality annotations detection. Extensive experiments show i) feasible; ii) model pre-trained on helps performance; iii) DVDET also cars. encourage more researchers investigate area, released related code.