Reproduce “Deep Ordinal Regression Network for Monocular Depth Estimation” on OpenResearch

About this paper

Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Batmanghelich, Dacheng Tao

Researchers developed DORN, a Deep Ordinal Regression Network for monocular depth estimation, by reformulating the task as an ordinal regression problem with spacing-increasing discretization. This approach achieved state-of-the-art performance on four major benchmarks, including a 1st prize win in the Robust Vision Challenge 2018, while improving accuracy and efficiency.

Key insights

• Re-framing monocular depth estimation as an ordinal regression problem, explicitly leveraging the ordered relationship between depth values, improves training robustness and accuracy compared to standard continuous regression.
• Spacing-Increasing Discretization (SID), which partitions depth in log-space, effectively addresses the non-uniform uncertainty in depth predictions by fine-graining nearby depths and coarsely grouping distant ones, thereby down-weighting errors for large depth values.
• High-resolution dense prediction can be achieved efficiently by maintaining feature map resolution using dilated convolutions and integrating a lightweight global context module, avoiding the need for complex multi-scale or deconvolutional architectures.

Summary by alphaXiv.