ZoeDepth: Zero-shot Transfer by Combining Relative and Metric Depth

Source: Dev.to

Overview

ZoeDepth predicts depth from a single image, handling both near and far objects accurately. It combines two learning strategies: one that preserves metric scale and another that captures relative shape.

Method

• Dual learning: The model is first trained on large, diverse datasets using a relative‑depth objective, then fine‑tuned with metric‑depth supervision to retain real‑world scale.
• Dynamic specialist selection: An internal selector chooses the most suitable “tiny specialist” for each input image, ensuring optimal performance across varied scenes.

Generalization and Performance

• Trained on many datasets, ZoeDepth generalizes well to unseen indoor and outdoor environments.
• Demonstrates strong zero‑shot depth estimation, outperforming prior methods without additional training.

Applications

• Photo‑editing effects that rely on depth cues.
• Space measurement tools for interior design or architecture.
• Robotics and navigation systems that need reliable depth perception.

Availability

The code and pretrained models are publicly released for community use.

ZoeDepth: Zero-shot Transfer by Combining Relative and Metric Depth