Deep Monocular Depth Estimation Based on Content and Contextual Features.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning methods for monocular depth estimation often struggle with low-texture areas and occlusions.
• Existing approaches primarily rely on RGB image content and structure, limiting accuracy.

Purpose of the Study:

• To propose a novel method for precise monocular depth map estimation using contextual semantic information.
• To enhance the accuracy and robustness of depth estimation by preserving depth discontinuities and object boundaries.

Main Methods:

• Leveraging a deep autoencoder network integrated with high-quality semantic features from HRNet-v2 semantic segmentation.
• Utilizing object localization and boundary information from semantic features to guide depth prediction.

Main Results:

• Achieved 85% accuracy on NYU Depth v2 and SUN RGB-D datasets.
• Outperformed state-of-the-art methods, reducing error metrics (Rel by 0.12, RMS by 0.523, log10 by 0.0527).
• Demonstrated superior performance in preserving object boundaries and detecting small structures.

Conclusions:

• The proposed semantic-feature-enhanced method significantly improves monocular depth estimation accuracy and robustness.
• Exploiting contextual semantic information is effective for overcoming limitations of textureless regions and occlusions in depth prediction.