Apple releases Depth Pro 2 for sharp, temporally consistent metric monocular depth
·2026.11.21 04:19
Key point
The model processes images and video with a single network to produce depth maps that are metrically accurate, sharp, and temporally consistent simultaneously.
Details
- Depth Pro 2 estimates dense metric depth for images and video at full resolution using a single scalable network.
- The approach overcomes the trade-off between temporal consistency and boundary sharpness found in prior monocular depth estimation methods.
- Released models define a new Pareto frontier in terms of metric accuracy, sharpness, temporal consistency, and runtime.
- Code and weights are available on GitHub for use with ordinary camera inputs.
Technical Approach
- Analyzes prior methods to identify that multi-frame methods favor consistency while single-frame methods favor sharpness.
- Trains a network that handles single photos, whole videos, or live camera feeds at the input's native resolution.
- Produces depth maps that are metrically accurate, sharp at object edges, and steady from frame to frame.
Availability
- Published in November 2026 as a computer vision research paper.
- Source code and model weights released via Apple's machine learning research repository.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.