Discard a bad calibration image or retake the pose

You captured an 18-image calibration set and the solver reports twice your usual error. One frame is obviously soft, or the corner detector half-missed the board, and the question is practical: throw that one image away, or redo the whole capture? In any tool that reports per-view reprojection error you discard the bad pose and recalibrate with the rest — no restart. This guide gives the thresholds that decide which pose goes, the culling procedure step by step, a worked example on an 18-pose set, and the capture checks that keep fixed-sequence robot routines from ever reaching a restart.

The short answer

Two cases, one decision each.

If your tool reports per-view reprojection error — OpenCV's calibrateCamera returns it, and most offline calibration utilities print it per image — discard the bad pose and recalibrate. Drop any view above about 1 px when the set's other views sit near 0.3 px, then re-read every per-view error, because each deletion refits the model for the survivors. Keep at least 10 distinct poses; an 18-image set loses two or three without any retake.

If your tool locks the pose sequence — several robot hand-eye wizards capture a fixed set of 15 to 18 positions with no per-image delete — one bad frame ends the run, and you recapture the set, about three minutes. There the win is prevention: check each pose before the sequence advances.

For the overall targets that average and max residuals should meet, read camera calibration residual values: average and max targets.

What makes one pose bad

In rough order of how often each one sends an image to the bin:

Focus spread. Tilting the board moves its far edge away from the lens. At 45°, a 100 mm board's far edge sits about 70 mm behind its near edge (100 × sin 45° = 71 mm). Close-range lenses hold far less depth of field than that, so the far corners soften while the center stays crisp. Keep tilts nearer 20–40° and stop the aperture down when the board is close.

Clipped highlights and glare. An overexposed white square bleeds into its neighbors and drags the detected corner inward, and a laminated or glossy target adds a specular band. Expose so the white squares sit near 230 of 255 rather than pegged at 255, and kill reflections with cross polarization to reduce surface glare.

Detection faults. The grid declared to the solver does not match the physical board — 9 × 6 told, 9 × 5 detected — the board runs partly out of frame, or the detector locks onto a reflection of the board. A mismatched pitch is the classic one; checkerboard calibration error: fix the grid pitch walks through the fix.

Weak set geometry. When most poses are near-frontal at one distance, the solver leans on the few tilted views, so error concentrates exactly on the informative poses. Before discarding a high-error tilted pose, count how many tilted poses would remain.

Motion blur. A board on the robot flange that is still settling when the frame is grabbed smears every corner. A global-shutter model such as the MV-CH120-11UC grabs a clean frame while the board is still settling:

The numbers that decide it

Four numbers, used together:

• Total RMS: 0.2–0.35 px from a rigid target and a sound lens; accept up to 0.5 px. A total above 1 px is a set-wide problem, not one image.
• Per-view error: open any view above twice the set median; discard a view above about 1 px when the median sits near 0.3 px.
• Pose count: 10 distinct poses minimum.
• Parameter stability: focal length within 0.5 % and principal point within 1 px of the previous run.

x_i is a detected corner in pixels, x_proj,i its reprojection through the fitted model, and n the corners in that image. The set RMS is the same sum over every corner of every kept image.

A low RMS is necessary but not sufficient — camera extrinsic calibration accuracy beyond reprojection error covers what the RMS number cannot tell you.

How to cull and retake one pose

1. Calibrate with all images. Record the total RMS and every per-view error.
2. Sort the per-view errors and open the worst image beside a mid-ranked one.
3. Judge the image. Soft corners, clipped whites or a missed corner grid: discard it. Sharp and clean: suspect set geometry or the lens model before the image.
4. Recalibrate with that one pose removed.
5. Re-read all per-view errors — they move after every refit — and repeat steps 2 to 4 one pose at a time.
6. Stop when the total RMS is at or under 0.5 px and the worst kept view at or under 1 px.
7. Compare focal length and principal point against the previous run. Inside 0.5 % and 1 px, the cull is sound.
8. Below 10 kept poses, capture replacements at a new tilt and distance before accepting anything.

Worked example: culling two poses from an 18-image set

Inputs: a 5 MP camera, 2448 × 2048 pixels at 3.45 µm each; a 200 mm wide field of view; a rigid 100 mm dot-grid board at varying tilt; 18 poses captured at tilts of 20–40°.

Scale first. One pixel on the part is 200 mm ÷ 2448 px = 0.082 mm, so every 0.1 px of reprojection error is 8 µm of image-plane disagreement.

Run A, all 18 poses: total RMS 0.48 px. Sixteen views sit between 0.21 and 0.44 px, pose 7 at 1.02 px, pose 14 at 1.58 px. Pose 14's error on the part is 1.58 px × 0.082 mm/px = 0.13 mm — larger than a 0.1 mm placement tolerance. Opened, pose 14 shows a glare band across the top two rows. Discard it.

Run B, 17 poses: total RMS 0.39 px. Pose 7 is now 0.87 px, still nearly three times the 0.30 px median. Opened, its far corners are soft from a steep tilt. Discard it.

Run C, 16 poses: total RMS 0.33 px, worst kept view 0.46 px. Stability: focal length moved from 7246 to 7258 px, a 0.17 % change; principal point moved 0.6 px.

Decision: accept Run C. Sixteen poses clear the 10-pose floor, both error thresholds are met, and the parameters barely moved — no retakes needed. Had the cull taken the set below 10 poses, the rule flips: capture two or three replacement poses first, then accept.

When you cannot delete a pose

Some robot-vision wizards and a few camera tools lock the sequence: 15 to 18 fixed positions, captured and accepted one by one, no delete. Frame 12 blurred means the run dies and you recapture — about three minutes. Buy your margin at capture time:

• Watch the live detection at each position. If the corners do not lock with the full count, or the board looks soft, reposition and recapture that position before the sequence advances.
• A frame that never arrives kills the run just as dead; software trigger image timeout: why one-shot capture misses frames lists the usual causes.
• In hand-eye calibration each image must stay paired with its robot pose — reading the robot TCP pose for each capture keeps that pairing intact when a position is retaken. For which positions to pick in the first place, see eye in hand calibration pose selection: robot positions that work.
• Only enter the cell with the robot in its safely reduced state, and keep the guarding live while you hold the target.

Checks before you blame the image

• All views uniformly between 0.5 and 0.9 px is not a bad-image signature; it is lens aberration, overall soft focus, a thin distortion model, or a target that is not flat. A paper print glued without backing adds error that no culling removes.
• Ambient light stealing contrast off the board shows up as uniformly weak corners; machine vision stray light: diagnose low image contrast covers the diagnosis.
• Dot grids avoid the corner-count ambiguity of chessboards, and monochrome images detect more cleanly than color — circular grid calibration with monochrome images. The LS-LXR-0029 board carries a 0.5 mm dot grid across a 76 x 76 mm pattern:

• A stand makes framing repeatable between culls and retakes. The LS-MNT-0040 reaches a 700 mm height on a 400 x 450 mm base:

Browse the full range of calibration targets if your board size does not match your field of view.

Checklist before you accept any calibration set

• Every kept pose: full corner count detected, sharp at all four corners, no clipped whites.
• Board covers a third to a half of the frame, tilts 20–40°, distances and frame positions varied.
• At least 10 kept poses, no two near-identical.
• Total RMS at or under 0.5 px; worst kept view at or under 1 px.
• Focal length and principal point stable within 0.5 % and 1 px across the final cull.
• One independent check: measure an object of known length in the calibrated image at two working distances.

If the hardware is what holds the error up, the camera finder and the system configurator are the quickest next steps, and the contact page puts an applications engineer on the question.