Monocular Depth Estimation for Autonomous Driving Based on Instance Clustering Guidance

A novel monocular depth estimator for autonomous driving, which produces reliable instance depths via instance clustering guidance, is proposed in this work. First, we extract multi-scale feature maps from a road scene and initialize depth clusters. Second, we update the depth clusters using the feature maps through transposed cross-attention. To guide the update process, we develop the instance clustering membership (ICM) loss, which employs an instance segmentation map. Third, we transfer the updated depth clusters to the feature map at the finest resolution, from which we produce the final depth map. Extensive experimental results show that the proposed algorithm yields competitive results to state-of-the-art techniques on the KITTI, Cityscapes-DVPS, and SemKITTI-DVPS datasets.