A Multi-Feature-Based Multi-Task Loss Approach for Enhanced Monocular Depth Estimation

This paper proposes a novel approach for monocular depth estimation that leverages deep networks and employs three types of features, namely depth map, point cloud, and virtual normal, to compute a multi-task loss function. Our experimental evaluations employ one cutting-edge monocular depth estimation model with distinct convolutional network architectures, trained and tested on a public indoor depth estimation NYU dataset. The experimental results demonstrate the consistent improvement in accuracy of depth estimation achieved by minimizing the proposed multi-task loss function, thus validating its effectiveness.