<p>Monocular 3D object detection is a challenging task of autonomous driving. Its performance is limited by its inherent depth ill-posed problem. To alleviate that problem, some methods exploit the depth information of LiDAR through knowledge distillation. However, existing methods with knowledge distillation usually apply strict alignment for feature distillation, and could lead to suboptimal performance because attributes of different modals may not be well aligned. To resolve this feature misalignment issue, we propose DGKD, a novel monocular 3D object detection network with Depth-Guided Knowledge Distillation to fully leverage the depth information from LiDAR. Specifically, for feature distillation, we propose global relation distillation to capture structural information of the feature maps for better knowledge transfer from the teacher network to the student network. For response distillation, we present Wasserstein Distance-based logit distillation to perform cross-category comparison, which is a promising alternative of commonly-used Kullback-Leibler Divergence (KL-Div) method. Extensive experiments on the KITTI dataset demonstrate that our DGKD achieves competitive performance.</p>