Progressive Hard-Mining Network for Monocular Depth Estimation.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Monocular depth estimation from RGB images is challenging due to limited prior knowledge.
• Existing methods often fail to preserve crucial cross-border details in depth maps, hindering performance.

Purpose of the Study:

• To propose a novel end-to-end progressive hard-mining network (PHN) framework for accurate monocular depth estimation.
• To address the limitation of preserving cross-border details in depth map generation.

Main Methods:

• Developed a progressive hard-mining network (PHN) framework.
• Introduced a hard-mining objective function and intra-scale/inter-scale refinement subnetworks.
• Designed a difficulty-aware refinement loss function to focus on error-prone regions.

Main Results:

• The PHN framework effectively localizes and refines hard-mining regions in depth maps.
• Intra-scale and inter-scale refinement blocks recursively recover and complement depth details.
• Evaluations on NYU Depth V2, KITTI, and Make3D datasets demonstrated superior performance over state-of-the-art methods.

Conclusions:

• The proposed PHN framework significantly enhances monocular depth estimation accuracy.
• The collaborative modules progressively reduce error propagation for improved depth map quality.
• The method shows strong potential for real-world applications requiring precise depth information.