MonoA2: Adaptive depth with augmented head for monocular 3D object detection

Abstract

Monocular 3D object detection is a hot direction due to its low cost and configuration simplicity. Achieving accurate instance depth prediction from monocular images is a challenging problem in monocular 3D object detection. Many existing methods perform instance depth prediction based on fixed rules, which are not flexible for various objects. Furthermore, these methods ignore the design of more discriminative task heads. To address these issues, we propose the MonoA${}^2$, which consists of the Adaptive Depth Module (ADM) and the Augmented Head Module (AHM). The ADM is used to achieve more accurate depth prediction by learning adaptive offsets to decouple the depth prediction from object center constraints. The AHM is proposed to obtain more discriminative task heads through task-aware attention and task-interaction attention. The task-aware attention can generate different weights adapted to different tasks and the task-interaction attention can guide depth tasks to interact with other tasks. Experimental results on the KITTI and Waymo datasets demonstrate the effectiveness of the proposed method. Our method achieves superior performance on the KITTI and Waymo benchmarks.