Shape-Aware Monocular 3D Object Detection

The 3D object detection is the key issue in autonomous driving system. This particularly challenging when only relies on a single perspective camera. anchor-free and keypoint-based models receive increasing attention recently due to their effectiveness simplicity. However, most of these methods are vulnerable occlusion truncation objects. In this paper, single-stage monocular model proposed. An instance-segmentation head integrated into training, which allows be aware visible shape target object. Therefore, largely avoids interference from irrelevant regions surrounding addition, we also reveal that popular IoU-based evaluation metrics, were originally designed for evaluating stereo or LiDAR-based methods, insensitive improvement achieved by algorithms. A novel metric, namely average depth similarity (ADS) proposed models. Our method outperforms comparison baseline terms both metrics while maintaining real-time efficiency.