UniDepthV2: Universal Monocular Metric Depth Estimation Made Simpler

Area of Science:

• Computer Vision
• 3D Reconstruction
• Machine Learning

Background:

• Monocular metric depth estimation (MMDE) is vital for 3D perception but current methods lack domain generalization.
• Existing MMDE models perform poorly on unseen data, limiting practical applications.

Purpose of the Study:

• To develop a universal monocular metric depth estimation (MMDE) model, UniDepthV2, that generalizes across diverse domains.
• To enable accurate 3D scene reconstruction from single images without domain-specific training.

Main Methods:

• UniDepthV2 employs a self-promptable camera module and a pseudo-spherical output representation to disentangle camera and depth features.
• Introduced a geometric invariance loss and an edge-guided loss for improved feature invariance and edge sharpness.
• Utilized a simplified, efficient architecture with an added uncertainty-level output.

Main Results:

• UniDepthV2 demonstrates superior zero-shot generalization across ten diverse depth datasets.
• The model achieves enhanced edge localization and sharpness in metric depth outputs.
• The uncertainty-level output provides confidence measures for downstream tasks.

Conclusions:

• UniDepthV2 offers a universal and flexible solution for monocular metric depth estimation, significantly improving domain generalization.
• The proposed methods enhance the accuracy, robustness, and applicability of single-image 3D reconstruction.