Improving Depth Gradient Continuity in Transformers: A Comparative Study on Monocular Depth Estimation with CNN

• DOI: 10.48550/arxiv.2308.08333
• Publication Date: 2023-01-01
• Authors: Jiawei Yao; Tong Wu; Xiaofeng Zhang
• Keywords: Monocular; Leverage (statistics)

Monocular depth estimation is an ongoing challenge in computer vision. Recent progress with Transformer models has demonstrated notable advantages over conventional CNNs this area. However, there's still a gap understanding how these prioritize different regions 2D images and affect performance. To explore the differences between Transformers CNNs, we employ sparse pixel approach to contrastively analyze distinctions two. Our findings suggest that while excel handling global context intricate textures, they lag behind preserving gradient continuity. further enhance performance of monocular estimation, propose Depth Gradient Refinement (DGR) module refines through high-order differentiation, feature fusion, recalibration. Additionally, leverage optimal transport theory, treating maps as spatial probability distributions, distance loss function optimize our model. Experimental results demonstrate integrated plug-and-play proposed without increasing complexity computational costs on both outdoor KITTI indoor NYU-Depth-v2 datasets. This research not only offers fresh insights into but also paves way for novel methodologies.