VGLD: Visually-guided language disambiguation for monocular depth scale recovery.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Monocular depth estimation is crucial for understanding 3D scenes from 2D images.
• Relative depth estimation lacks absolute scale, limiting its use in real-world applications.
• Inferring metric scale from textual descriptions is challenging due to natural language ambiguity.

Purpose of the Study:

• To develop a method for visually grounded language disambiguation to improve metric depth estimation.
• To address the limitations of purely language-based approaches in recovering accurate metric depth.
• To create a universal alignment module for enhancing depth estimation models.

Main Methods:

• Introduced Visually-Guided Language Disambiguation (VGLD) to leverage visual semantics for disambiguation.
• VGLD predicts global linear transformation parameters to align relative depth with metric scale.
• Evaluated VGLD on MiDaS and Depth Anything models using NYU Depth V2 and KITTI datasets.

Main Results:

• VGLD effectively mitigates language-induced scale bias in metric depth estimation.
• Demonstrated significant improvements in metric depth accuracy across benchmark datasets.
• VGLD functions as a lightweight, universal alignment module, maintaining strong zero-shot performance.

Conclusions:

• Visually-Guided Language Disambiguation offers a robust solution for accurate metric depth estimation.
• Integrating visual semantics enhances the reliability of text-guided depth scale recovery.
• VGLD shows promise as a versatile tool for improving various depth estimation models.