Reproduce “HR-Depth: High Resolution Self-Supervised Monocular Depth Estimation” on OpenResearch

About this paper

Xiaoyang Lyu, Liang Liu, Mengmeng Wang, Xin Kong, Lina Liu, Yong Liu, Xinxin Chen, Yi Yuan

HR-Depth introduces a self-supervised monocular depth estimation network that significantly improves depth accuracy, particularly at object boundaries, by effectively leveraging high-resolution inputs. Its lightweight variant, Lite-HR-Depth, achieves performance comparable to larger state-of-the-art models with only 3.1 million parameters, enabling efficient deployment on resource-constrained platforms.

Key insights

• The primary limitation of prior MDE methods in leveraging high-resolution inputs stemmed from inaccurate depth estimation in 'large gradient regions' (object boundaries) where standard interpolation methods perform poorly.
• Gradually reducing the semantic gap through dense skip-connections is crucial for effectively fusing multi-scale features, enabling the network to produce sharper and more accurate depth predictions, especially at edges.
• The fSE module demonstrates that sophisticated, channel-reweighting fusion mechanisms can simultaneously improve accuracy and significantly reduce model parameters compared to standard convolutional fusion blocks.

Summary by alphaXiv.