Deep Ordinal Regression Network for Monocular Depth Estimation.

Area of Science:

• Computer Vision
• Machine Learning
• 3D Scene Understanding

Background:

• Monocular depth estimation is vital for 3D scene geometry but is an ill-posed problem.
• Current deep convolutional neural network (DCNN) methods use regression with mean squared error, leading to slow convergence and suboptimal solutions.
• Existing networks often produce low-resolution feature maps due to spatial pooling, requiring complex additions for high-resolution outputs.

Purpose of the Study:

• To address the limitations of existing monocular depth estimation methods, including slow convergence and low-resolution outputs.
• To introduce a new strategy for discretizing depth and reformulating the learning problem.
• To develop a more efficient and accurate deep learning model for depth estimation.

Main Methods:

• Introduced a spacing-increasing discretization (SID) strategy to discretize depth values.
• Recast depth network learning as an ordinal regression problem, trained with an ordinal regression loss.
• Employed a multi-scale network structure to avoid excessive spatial pooling and capture information at various scales.

Main Results:

• The proposed deep ordinal regression network (DORN) achieved significantly higher accuracy and faster convergence compared to existing methods.
• The multi-scale architecture effectively captured multi-scale information in parallel without unnecessary spatial pooling.
• State-of-the-art results were obtained on the KITTI, Make3D, and NYU Depth v2 benchmarks.

Conclusions:

• The ordinal regression approach with SID is highly effective for monocular depth estimation.
• The DORN model offers a more efficient and accurate solution for generating high-resolution depth maps.
• This method represents a significant advancement in monocular depth estimation research.