Frequency-Geometry-Guided Network for Depth Map Super-Resolution.

Area of Science:

• Computer Vision
• Image Processing
• Deep Learning

Background:

• Depth super-resolution aims to generate high-resolution (HR) depth maps from low-resolution (LR) inputs using HR RGB guidance.
• Existing methods struggle with RGB edges not aligning with depth discontinuities, leading to texture copying and geometric inconsistencies.

Purpose of the Study:

• To propose Frequency-Geometry-Guided Network (FGGNet), a novel framework for RGB-guided depth map super-resolution.
• To improve geometric consistency and reduce texture copying in reconstructed depth maps.

Main Methods:

• FGGNet employs a spatial-frequency fusion framework.
• Key components include Multi-branch RGB-guided Convolution (MRGConv) for enhanced RGB representations and a Geometry Prior-guided Fusion Module (GPFM) for filtering inconsistent RGB responses.
• Radial complex spectral loss (RCSL) is used to emphasize high-frequency components in the complex spectral domain.

Main Results:

• FGGNet demonstrated competitive or superior reconstruction accuracy on NYU v2, Middlebury, Lu, and RGB-D-D datasets.
• Achieved significant Root Mean Square Error (RMSE) reductions under ×16 magnification: 13.7% (NYU v2), 22.8% (Middlebury), 18.5% (Lu), and 11.4% (RGB-D-D) compared to state-of-the-art methods.
• Validated effectiveness across synthetic and real-world degradation settings.

Conclusions:

• The proposed FGGNet effectively enhances RGB-guided depth map super-resolution.
• Combining geometric prior filtering with frequency-domain supervision is crucial for reliable depth reconstruction.