NDDepth: Normal-Distance Assisted Monocular Depth Estimation and Completion.

Area of Science:

• Computer Vision
• Deep Learning
• 3D Scene Understanding

Background:

• Monocular depth estimation and completion are crucial computer vision tasks with broad applications.
• Existing methods often directly estimate depth, which can be challenging with sparse or incomplete data.

Purpose of the Study:

• To develop novel deep learning frameworks for monocular depth estimation and completion.
• To leverage physics (geometry)-driven principles by assuming 3D scenes are composed of piece-wise planes.

Main Methods:

• Proposed a framework that estimates surface normal and plane-to-origin distance maps as intermediate representations.
• Developed a normal-distance head for pixel-level output and a plane-aware consistency constraint for regularization.
• Integrated an additional depth head to enhance robustness and transformed intermediate outputs into depth maps.

Main Results:

• The proposed method demonstrated superior performance compared to state-of-the-art competitors.
• Experiments were conducted on widely recognized datasets: NYU-Depth-v2, KITTI, and SUN RGB-D.
• The physics-driven approach proved effective for both depth estimation and completion tasks.

Conclusions:

• The novel deep learning frameworks offer a robust and high-performing solution for monocular depth estimation and completion.
• The assumption of piece-wise planar scenes and the use of intermediate surface normal and distance maps are key to the method's success.
• This work advances the field of 3D scene understanding from monocular images.