A perception team chasing a persistent 3-pixel bias in its object centroids spent two weeks retraining the detector. The model was never the problem. The camera intrinsic parameters were wrong, and every projection in the pipeline inherited the same systematic warp — uniformly, quietly, and in a way no amount of retraining could fix.

That misattribution is common, and it is expensive. Camera intrinsic parameters describe how a specific lens-and-sensor combination maps points in 3D space onto the 2D image plane. When they are wrong, the error is not random noise you can average away — it is a fixed geometric distortion baked into every frame the camera produces. The point of this article is to make the individual parameters concrete, so that when a projection looks wrong you know whether to suspect the calibration or the model.

What do focal length, principal point, and distortion coefficients each represent physically?

Intrinsics are usually packed into a 3×3 matrix (often written K) plus a short vector of distortion coefficients. Teams that treat that matrix as an opaque vendor spec lose the ability to reason about it. Broken open, it is four physical quantities and a curvature correction.

Focal length (fx, fy, in pixels) sets the scale of the pinhole projection — how many pixels a real-world angle subtends. A longer focal length magnifies the scene; a shorter one widens the field of view. Because it is expressed in pixels, it couples the lens’s optical focal length with the sensor’s pixel pitch. fx and fy differ only when pixels are non-square, which is rare in modern automotive sensors but not something to assume away.

Principal point (cx, cy, in pixels) is where the optical axis pierces the image plane — the true center of projection, which is almost never the exact geometric center of the sensor. A principal point offset of a few pixels is normal; getting it wrong shifts the entire projection.

Distortion coefficients correct the lens’s departure from the ideal pinhole model. Radial coefficients (k1, k2, k3) handle the barrel or pincushion curvature that bends straight lines near the image edges; tangential coefficients (p1, p2) handle the slight misalignment between lens and sensor. Wide-angle automotive lenses distort heavily, and an uncorrected distortion field is a projection error that grows toward the periphery.

Here is the distinction that most black-box treatments miss: focal length and principal point errors distort the projection uniformly across the whole image, while distortion-coefficient errors concentrate at the edges. That difference is a diagnostic signal, and we return to it below.

Quick-answer: what each intrinsic controls

Parameter	Symbol	Physical meaning	Error signature
Focal length	fx, fy	Projection scale (angle → pixels)	Uniform scale/magnification error everywhere
Principal point	cx, cy	Where the optical axis hits the sensor	Uniform positional shift everywhere
Radial distortion	k1, k2, k3	Barrel/pincushion lens curvature	Error grows toward image edges
Tangential distortion	p1, p2	Lens-sensor misalignment	Asymmetric edge error

How are camera intrinsic parameters estimated during calibration?

Intrinsics are recovered empirically, not read off a datasheet. The standard procedure — implemented in OpenCV’s calibrateCamera, MATLAB’s Camera Calibrator, and the calibration routines inside ROS — shows the camera a target of known geometry (a checkerboard or a ChArUco board) from many angles. Each captured view gives the solver a set of correspondences between known 3D board points and their observed 2D pixel locations. The optimizer then finds the intrinsic matrix and distortion coefficients that best explain all those correspondences at once.

The quality metric that falls out of this is reprojection error: take the estimated intrinsics, project the known 3D points back onto the image, and measure the pixel distance to where they were actually observed. In configurations we have worked with, a sub-pixel mean reprojection residual (roughly under one pixel) is the threshold below which a calibration is usually trustworthy for perception work — though this is an observed working heuristic, not a universal benchmark, and the acceptable bound depends on sensor resolution and the downstream task.

Two practical failure modes recur. First, insufficient angular coverage: if every calibration view is roughly frontal, the solver cannot separate focal length from board distance, and the estimate is degenerate even when reprojection error looks low. Second, a target that is too small in the frame leaves the distortion coefficients under-constrained precisely where distortion matters most — near the edges. Getting the estimation right is the same class of discipline we describe in the companion piece on why intrinsic traceability matters, where the emphasis is on recording and versioning the result rather than on the parameters themselves.

Why do intrinsic parameters stay fixed unless the lens or sensor changes?

This is the property that makes intrinsics both easier and more dangerous than extrinsics. Extrinsics — the camera’s position and orientation relative to the vehicle — can drift with vibration, thermal cycling, or a minor knock, which is why they belong in a monitoring regime, a point developed in our discussion of what camera extrinsics mean in practice. Intrinsics do not drift in the same way. Focal length, principal point, and distortion are properties of the optics and the sensor. Absent a lens swap, a mechanical shift in the lens mount, or a sensor replacement, they are constant for the life of the module.

The consequence cuts both ways. On one hand, you calibrate intrinsics once per hardware configuration and reuse the result — no continuous re-estimation needed. On the other hand, a bad intrinsic calibration is a persistent, systematic error rather than a drift event. It does not announce itself with a gradual degradation you might catch in monitoring dashboards. It is wrong from the first frame and stays exactly as wrong until someone re-calibrates. That stability is precisely what makes it easy to misattribute to the perception model, because it never behaves like a fault — it behaves like a consistent, reproducible offset.

How does a wrong intrinsic estimate distort projection across the whole image?

Consider a worked example, with the assumptions stated explicitly. Suppose the true focal length fx is 1000 pixels but the calibration recovered 1010 — a 1% error. Every 3D point projected through that matrix lands about 1% too far from the principal point. An object whose correct image position is 400 pixels from center appears at roughly 404; one at 600 pixels appears near 606. The error is not constant across the frame — it scales with distance from the principal point, so it looks like a subtle radial stretch.

Now suppose instead the principal point cx is off by 4 pixels. Every projected point shifts by that same 4 pixels in the same direction, everywhere in the image. This is the uniform bias that sent the team in the opening paragraph chasing their detector.

The reason this matters for downstream perception is that the intrinsic matrix sits at the seam between the 2D image and the 3D world. A detector that outputs pixel-perfect bounding boxes still feeds those boxes through the intrinsics to estimate distance, heading, or bird’s-eye-view position. If the intrinsics are wrong, the pixels are right and the 3D estimate is wrong — a failure mode that survives every image-space benchmark, in exactly the way we catalogue in the perception failure modes that survive benchmarks. The model scores well on detection metrics and still places objects in the wrong physical location.

How can a fleet-perception team tell whether an error comes from bad intrinsics or the model?

This is the diagnostic that turns the theory into an audit. The signatures are distinguishable if you know what to look for.

Diagnostic checklist: intrinsics or model?

• Is the error a consistent geometric offset across frames, scenes, and object classes? A model error varies with scene content and object appearance; an intrinsic error is the same warp regardless of what is in the frame. Consistency points to calibration.
• Does the error scale with distance from image center? A focal-length error stretches radially; a distortion-coefficient error concentrates at the edges. A model error has no reason to respect the image geometry this cleanly.
• Does re-running against a calibration target reproduce a high reprojection residual? This is the definitive test. Reshoot a checkerboard, run calibrateCamera, and measure the residual. A sub-pixel residual exonerates the intrinsics; a large one indicts them directly (a benchmark-class measurement, because the target geometry is known and the test is reproducible).
• Is the error present in image space or only after 3D projection? If the 2D detections are correct but the 3D positions are biased, suspect the projection math — intrinsics or extrinsics — before the model.

Because intrinsic calibration produces this kind of reproducible, target-based evidence, it belongs in the same silent-failure surface a robustness audit checks before blaming the model. Verifying intrinsics is one line item in TechnoLynx’s Production AI Monitoring Harness, and it sits alongside the broader computer vision work where the question “is it the model or the calibration?” comes up on nearly every fleet-perception engagement. We treat that question as an audit step, not a debate — the reprojection test settles it.

FAQ

How does camera intrinsic parameters actually work?

Camera intrinsic parameters describe how a specific lens-and-sensor combination maps 3D points onto the 2D image plane, via a pinhole projection (focal length and principal point) plus a distortion correction. In practice they are recovered by calibration against a known target and reused for the life of the hardware configuration, because a wrong estimate systematically warps every projection the camera produces.

What do focal length, principal point, and distortion coefficients each represent physically?

Focal length (fx, fy) sets the projection scale — how many pixels a real-world angle subtends, coupling optical focal length with sensor pixel pitch. Principal point (cx, cy) is where the optical axis meets the sensor, which is almost never the exact geometric center. Distortion coefficients (k1–k3 radial, p1–p2 tangential) correct the lens’s departure from the ideal pinhole model, an effect that grows toward the image edges.

How are camera intrinsic parameters estimated during calibration?

You show the camera a target of known geometry — a checkerboard or ChArUco board — from many angles, and a solver (as in OpenCV’s calibrateCamera) finds the intrinsic matrix and distortion coefficients that best explain all the observed 3D-to-2D correspondences. The quality metric is reprojection error, the pixel distance between projected known points and their observed positions; a sub-pixel mean residual is a common working threshold for trustworthy calibration.

Why do intrinsic parameters stay fixed unless the lens or sensor changes, unlike extrinsics?

Focal length, principal point, and distortion are physical properties of the optics and sensor, so they do not drift with vibration or thermal cycling the way extrinsics do. Absent a lens swap, a mount shift, or a sensor replacement, they are constant — which means you calibrate once per hardware configuration, but a bad calibration is a persistent systematic error rather than a drift event you might catch in monitoring.

How does a wrong intrinsic estimate systematically distort 3D-to-2D projection across the whole image?

A focal-length error scales every projected point’s distance from the principal point — a 1% focal-length error stretches positions radially by roughly 1%. A principal-point error shifts every point uniformly in the same direction. Either way the distortion is baked into the whole image, so pixel-perfect detections still yield wrong 3D positions.

How can a fleet-perception team tell whether a projection error comes from bad intrinsics rather than the perception model?

Check whether the error is a consistent geometric offset across scenes and object classes, whether it scales cleanly with distance from image center, and — definitively — whether re-running a calibration target reproduces a high reprojection residual. A model error varies with scene content and has no reason to respect image geometry; an intrinsic error is the same reproducible warp regardless of what is in the frame.

The habit worth building is treating “the projection looks wrong” as a hypothesis to test, not a verdict on the model. Reshoot the target, measure the residual, and let the geometry answer. The teams that get burned are the ones that assume the black box was right and spend their weeks in the wrong place.