One snapshot reveals your camera’s hidden geometry

What it does GeoCalib is a research implementation (ECCV 2024) that reverse-engineers a camera’s internal parameters and gravity direction from a single photograph. It exposes a lightweight Python package, geocalib, that runs inference with minimal dependencies, while the heavier training and evaluation machinery lives in a companion library, siclib.

The interesting bit Instead of regressing parameters end to end with a neural network alone, GeoCalib folds deep learning into a geometric optimization pipeline. It also supports partial calibration: hand it a known focal length or gravity vector and it will refine the rest, or feed it multiple images from the same camera to tighten estimates under shared intrinsics.

Key highlights

• Supports several camera models out of the box, from vanilla pinhole to fisheye (simple_divisional) and radial distortion.
• Can calibrate with priors when either intrinsics or gravity are already known.
• Pools information across multiple images captured by the same camera.
• Published benchmarks on LaMAR, MegaDepth, and TartanAir show it outperforming DeepCalib, ParamNet, and UVP on roll, pitch, and field-of-view accuracy.
• Minimal-dependency inference package suitable for dropping into existing pipelines.

Caveats

• The principal point is hard-coded to the image center and never optimized.
• Distorted lenses require loading a separate weights="distorted" checkpoint; the default checkpoint expects pinhole images.
• Full training code and the OpenPano dataset are managed through the separate siclib library, not the core geocalib package.

Verdict Worth a look if you are building photogrammetry, AR, or robotics pipelines that need to auto-calibrate unknown cameras from sparse imagery. Skip it if you already have a calibrated rig or need to optimize the principal point.