Metric depth from one image, now left to the community

What it does

ZoeDepth is a PyTorch implementation that estimates metric depth—real-world physical distances—from a single RGB image. It ships three pretrained variants: ZoeD_N trained on NYU-Depth-v2, ZoeD_K trained on KITTI, and a multi-headed ZoeD_NK model. You can hand it a PIL image, a torch tensor, or a remote URL, and get back depth maps as numpy arrays, 16-bit PIL images, or tensors. The repo also includes training and evaluation scripts for both single-dataset and mixed-dataset setups.

The interesting bit

Most monocular depth estimators only produce relative up-to-scale depth; ZoeDepth attempts to recover true metric scale by combining relative and metric signals, as the paper title suggests. The codebase reflects this architecturally: ZoeD_N and ZoeD_K use a single metric head each, while ZoeD_NK uses multiple heads to handle both datasets at once. The README never explains how the fusion actually works, so the mechanism remains in the paper.

Key highlights

• Three pretrained variants: ZoeD_N (NYU-Depth-v2), ZoeD_K (KITTI), and multi-headed ZoeD_NK
• Inference accepts PIL images, torch tensors, or URLs; outputs numpy, 16-bit PIL, or tensors
• Loads via torch hub or local paths without manual weight hunting
• Training scripts for single-dataset (train_mono.py) and mixed-dataset (train_mix.py) workflows
• Gradio demo available locally and on HuggingFace Spaces

Caveats

• Intel has formally abandoned the project: no maintenance, bug fixes, patches, or updates accepted
• The README is essentially a quickstart guide; the underlying method is undocumented outside the paper

Verdict

A solid starting point if you need off-the-shelf metric depth inference and can tolerate unmaintained code. Skip it if you need active support, detailed documentation, or a living upstream.