Infrared Monocular Depth Estimation Based on Radiation Field Gradient Guidance and Semantic Priors in HSV Space

Area of Science:

• Computer Vision
• Computational Imaging
• Infrared Imaging

Background:

• Monocular depth estimation (MDE) using RGB images struggles with illumination variations.
• Infrared images offer illumination-invariant properties beneficial for depth estimation.

Purpose of the Study:

• Develop an end-to-end framework for accurate MDE using infrared images.
• Enhance structural learning and edge-aware mechanisms for improved depth prediction.

Main Methods:

• A multi-task UNet architecture for gradient extraction, semantic segmentation, and texture reconstruction from infrared RAW images.
• Incorporation of a Radiation Field Gradient Guidance (RGG) module for edge-aware attention.
• Mapping image features to HSV color space channels (S, H, V) and converting to RGB for the depth network.
• Introduction of a sky mask loss to handle ambiguous sky regions.

Main Results:

• Experimental validation on a custom infrared dataset demonstrated high accuracy.
• Achieved a delta-1 (δ1) accuracy score of 0.976.
• The proposed method significantly enhances depth estimation performance for infrared imagery.

Conclusions:

• Leveraging infrared images and the HSV color space improves MDE robustness to illumination changes.
• Radiation field gradient guidance and semantic priors are crucial for accurate depth estimation from infrared data.
• The developed framework offers a promising solution for MDE in challenging lighting conditions.