GFI-Net: Global Feature Interaction Network for Monocular Depth Estimation.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Monocular depth estimation recovers scene distances from single images.
• Existing methods struggle with accuracy, detail localization, and depth discontinuity.
• These limitations hinder real-world applications of depth estimation.

Purpose of the Study:

• To propose a novel network, the Global Feature Interaction Network (GFI-Net), for enhanced monocular depth estimation.
• To address limitations in accuracy, detail localization, and depth continuity.
• To leverage global geometric features for improved depth prediction.

Main Methods:

• Designed a Global Interactive Attention mechanism to capture feature graph interactions.
• Utilized a Transformer encoder to minimize coding losses.
• Implemented a local-global feature fusion module for detailed area representation.

Main Results:

• GFI-Net achieved state-of-the-art performance on NYU-Depth-v2 and KITTI datasets.
• Demonstrated significant improvements in full detail recovery.
• Showcased enhanced depth continuation on planes parallel to the camera.

Conclusions:

• GFI-Net effectively improves monocular depth estimation by integrating global geometric context.
• The proposed attention mechanism and fusion module are key to enhanced performance.
• The model offers a robust solution for accurate and detailed depth map generation.