Densely Connecting Depth Maps for Monocular Depth Estimation

Jinqing Zhang, Haosong Yue, Xingming Wu, Weihai Chen, Changyun Wen

Abstract: Predicting depth map from a single RGB image is beneficial for many three-dimensional applications. Although recent monocular depth estimation methods have achieved impressive accuracy, the preference on high-level features or low-level features prevents them from balancing sharpness and fidelity of depth maps. In this work, we propose a dense connection mechanism that connects diverse sub-depth maps produced by the sub-predictors to the final depth map to contribute information from features at different levels. Besides, two kinds of diversity enhancement devices are proposed to increase the number and diversity of the sub-depth maps collected by the dense connection mechanism. Experimental results on KITTI and NYU Depth V2 datasets shows that, by fusing the dense connection mechanism and diversity enhancement devices, our proposed method achieves state-of-the-art accuracy and predicts sharp depth maps that restore reliable object structures.