IGAF: Incremental Guided Attention Fusion for Depth Super-Resolution.

Area of Science:

• Computer Vision
• Robotics
• Sensor Fusion

Background:

• Accurate depth estimation is vital for robotics, navigation, and medical imaging.
• Conventional depth sensors yield low-resolution (LR) depth maps, limiting detailed scene perception.
• Enhancing LR depth maps to high-resolution (HR) using structured inputs like RGB images is essential.

Purpose of the Study:

• To propose a novel sensor fusion methodology for guided depth super-resolution (GDSR).
• To develop an Incremental guided attention fusion (IGAF) module for effective feature fusion.
• To create a robust super-resolution model for generating detailed HR depth maps from LR inputs.

Main Methods:

• Developed a novel sensor fusion methodology for guided depth super-resolution (GDSR).
• Introduced the Incremental guided attention fusion (IGAF) module to fuse RGB image and LR depth map features.
• Built and evaluated a super-resolution model using the IGAF module on benchmark datasets.

Main Results:

• The proposed GDSR model with IGAF achieved state-of-the-art results on the NYU v2 dataset for ×4, ×8, and ×16 upsampling.
• The model outperformed all baseline models in a zero-shot setting on Middlebury, Lu, and RGB-D-D datasets.
• Demonstrated accurate HR depth map generation through effective fusion of LR depth and HR image data.

Conclusions:

• The IGAF module effectively fuses features from RGB images and LR depth maps for accurate HR depth estimation.
• The proposed GDSR methodology provides a robust solution for enhancing depth map resolution.
• The approach achieves superior performance across multiple datasets, highlighting its generalizability and effectiveness.