Enhanced Monocular Depth Estimation: A CNN Integrating Semantic Segmentation Embedding And Vanishing Point Detection

This paper addresses the challenge of generating a depth map from a single RGB image of a specific environment. We have developed a unique approach that adapts semantic segmentation embedding layers, commonly utilized in NLP, and incorporates Manhattan frame vanishing point priors into a monocular depth estimation neural ne twork. Our me thod demonstrates robust performance on the NYU Depth V2 dataset and is more lightweight compared to previous methods, offering potential advantages for downstream tasks such as real-time 3D reconstruction and perception. Our codes and pretarined models are available.