Why I built realtime-calib: real-time, headless multi-camera calibration

Hans Bourgeois — Wed, 15 Jul 2026 13:53:49 +0000

I'm a freelance computer-vision developer with a PhD in human-movement science, and I build applied technology for health, physical activity and sports performance. This is the story of why I ended up writing my own multi-camera calibration tool — and why I open-sourced it.

The dependency I couldn't avoid

A lot of my client work rests on the same unglamorous first step: calibrating a rig of USB cameras. Before you can reconstruct a movement in 3D, track a joint or line up several viewpoints, every camera needs its intrinsics (focal length, lens distortion) and its extrinsics (its 6-DoF position and orientation in one shared coordinate frame). Get calibration wrong and everything downstream is wrong — so it kept coming back, project after project, as the quiet dependency I couldn't skip.

Where the existing tools rubbed

The best open-source reference I found was Caliscope. Its calibration math is
genuinely good — good enough that realtime-calib reimplements its logic rather than reinventing it. But the workflow around it kept getting in my way, in three recurring ways:

Record first, calibrate later. Each session meant pre-recording every camera (for me, through OBS and a sync plugin), then calibrating offline. That step was fragile: it crashed, and it sometimes lost frame sync silently, forcing manual video re-editing before I could even start.
No headless path. Some of my clients run on headless Linux VMs. Calibrating there meant standing up an extra machine with a desktop, passing the cameras through, and driving a GUI — a lot of ceremony just to get a calibration.
Export mismatch. The output convention didn't always match the target project's coordinate system, so results needed hand-conversion — exactly the kind of 3D-math bookkeeping that's easy to get subtly wrong.

None of this is Caliscope's fault; it wasn't built for my constraints. But together, these frictions made calibration slower and more error-prone than it needed to be.

What I actually wanted

Not a different calculator — the same solid math, with a different workflow around it:

calibrate live, in one pass, with no separate recording step;
run it on the camera host, even headless, and drive it from whatever device I had in hand — laptop, tablet or phone;
and get exports that already match the engine or tool the project targets.

What realtime-calib does

So I built realtime-calib. In one line: real-time, multi-camera calibration you drive from any device.

One pass, live. Capture, board detection, quality feedback and the solve happen in a single flow — what you see is what gets calibrated, no pre-recording.
Headless, any device. The service runs in Docker on the machine the cameras are plugged into — no desktop on that host — and you drive everything from a browser on any device on the local network. (Here's the architecture.)
Local, private, CPU-only. No cloud, no GPU; camera streams never leave your network.
Exports that fit. One calibration, written to the convention your target actually uses: Caliscope-compatible TOML, or engine-ready JSON with the correct axes and handedness for Unity, Unreal, Blender, three.js and ROS. The dangerous axis-remap math is done for you.

If you already use Caliscope, the TOML output keeps its semantics, so your existing pipelines keep working — I wrote a full side-by-side comparison if you're weighing the two.

Why open source

I could have kept this internal. Two things changed my mind.

First, multi-camera setups are everywhere now — robotics rigs, motion-capture rooms, volumetric capture, photogrammetry, production lines. As robotics and computer vision keep spreading, the need to calibrate several cameras into one coordinate frame only grows. A friction-free, self-hostable tool seemed worth sharing.

Second, calibration is a shared problem. The value isn't in hoarding it — it's in making it reliable for everyone who hits the same wall I did. So realtime-calib is open source under AGPL-3.0 (with a commercial option for proprietary products).

Try it — and tell me what's missing

realtime-calib is still early, and that's exactly why I'm writing this. What I want most right now is people to try it on their own rigs and tell me what breaks or what's missing, so it can grow from one freelancer's need into something genuinely useful.

If you set up multi-camera rigs — for mocap, robotics, volumetric or photogrammetry — I'd love your feedback.

Get started: the docs
Code & issues: GitHub

Transparency & acknowledgements: inspired by
Caliscope, created by Mac Prible. I use Claude Code (Opus 4.8) to assist me in writing the code.