Beyond Checkerboards – Bharadwaj & Lichti, IJCV 2026

Abstract

Imagery is fundamental to modern scientific research, making robust intrinsic camera calibration indispensable for accurate visual inference. The checkerboard-based calibration method has long been favored for its simplicity, and is widely used even in mission-critical computer vision pipelines. However, its limitations in modeling high-precision camera geometry compromise downstream performance. We revisit intrinsic calibration through the lens of photogrammetric self-calibration (PSC) and show it consistently outperforms checkerboard methods across 3D reconstruction, SfM, visual SLAM, and novel view synthesis.

Core finding: Reprojection error — the universal calibration quality metric — measures how well a model fits 2D image observations. It does not measure 3D accuracy. Checkerboard calibration minimizes reprojection error while leaving residual systematic distortions that compound into order-of-magnitude errors in reconstruction, localization, and synthesis.

Method

PSC uses a 3D target field with convergent imaging geometry and self-calibrating bundle adjustment to jointly refine interior orientation parameters (IOPs), exterior orientation parameters (EOPs), and target coordinates. The distortion model is selected by iteratively adding radial terms (K₃ → K₇) until all systematic residual patterns vanish.

Camera Model Selection — Radial Distortion Residuals

Residuals are plotted against radial distance after each bundle adjustment. Structured "bow-tie" patterns mean the model is insufficient. Adding K₇ eliminates all systematic error.

K₃

Strong systematic pattern

K₅

Pattern still present

K₆

Nearly eliminated

K₇ ✓

No systematic error — selected

SfM 3D Reconstruction

COLMAP + GloMap reconstruct two scenes with fixed intrinsics. Ground truth from a terrestrial LiDAR scanner. PSC reduces MAE by 5.5× (0.209 m → 0.038 m). Errors compound dramatically at outdoor scale — blurring text, distorting geometry, and curving straight façade lines.

Calibration	MAE (m)	RMSE (m)	Std (m)
Checkerboard CF	0.209	0.236	0.110
PSC (ours)	0.038	0.043	0.020

Errors vs. surveyed LiDAR ground truth.

Calibration Laboratory — Indoor

PSC produces tight clusters around ground-control points. CF leaves a visible spread — residual distortion the closed-form solver cannot eliminate.

PSCCF ↔ Drag

Top view — GCP cluster tightness

PSCCF ↔ Drag

Side view — point cloud noise

CCIT Building Facade — Large-Scale Outdoor

Drag each slider to compare. PSC preserves sharp text, correct bicycle geometry, and straight vertical lines. CF curves façade lines increasingly with height.

PSCCF ↔ Drag

View 1 — text legibility

PSCCF ↔ Drag

View 2 — bicycle geometry

PSCCF ↔ Drag

View 3 — façade line curvature

Visual Odometry & SLAM

Monocular ORB-SLAM3 on 9 sequences across the University of Calgary campus. Ground truth from FAST-LIO2 LiDAR-inertial odometry, averaged over 10 runs. PSC gains are largest in long outdoor and loop-closure scenarios where residual distortion accumulates into drift.

PSC (Bundle Adjustment) Checkerboard CF Ground Truth (FAST-LIO2)

Library Outdoor

Outdoor · Cart ~120 m

CF: 16.72 m→PSC: 1.21 m13.8×

Kinesiology Loop

Outdoor · Handheld ~35 m

CF: 6.91 m→PSC: 0.28 m24.7×

$Chem-Math$

Chem-Math Hallway

Indoor · Cart ~80 m

CF: 3.73 m→PSC: 0.43 m8.7×

Science-B Hallway

Indoor · Cart ~80 m

CF: 4.39 m→PSC: 0.52 m8.4×

Olympic Arch

Outdoor · Handheld ~20 m

CF: 0.39 m→PSC: 0.19 m2.1×

Mural

Indoor · Small Scale

CF: 0.20 m→PSC: 0.08 m2.5×

Novel View Synthesis

Nerfstudio Nerfacto (200K steps, no IOP refinement). Each scene card shows a draggable slider between PSC and CF renders. CF exhibits clouding, blurred text, and curved geometry — residual distortion the checkerboard solver failed to remove.

Calibration Laboratory

+5.4 dB

PSNR gain

+0.209

SSIM gain

3.2×

LPIPS ↓

PSCCF ↔ Drag

View 1 — pipes & clouding artifacts

PSCCF ↔ Drag

View 2 — text sharpness

PSC PSNR 25.46 · SSIM 0.914 · LPIPS 0.127 | CF PSNR 20.02 · SSIM 0.705 · LPIPS 0.411

Dino (Outdoor · Largest PSNR gap)

+6.8 dB

PSNR gain

+0.293

SSIM gain

3.1×

LPIPS ↓

PSCCF ↔ Drag

PSCCF ↔ Drag

PSC PSNR 23.20 · SSIM 0.861 · LPIPS 0.129 | CF PSNR 16.41 · SSIM 0.568 · LPIPS 0.396

Eleutheria

+6.4 dB

PSNR

1.7×

LPIPS ↓

PSCCF ↔ Drag

PSC 24.11 dB · 0.807 SSIM | CF 17.67 dB · 0.634 SSIM

Olympic Arch

+6.5 dB

PSNR

3.8×

LPIPS ↓

PSCCF ↔ Drag

PSC 24.01 dB · 0.913 SSIM · LPIPS 0.089 | CF 17.50 dB · 0.693 SSIM · LPIPS 0.341

NeRF Point Cloud — Calibration Lab

PSC — Denser geometry CF — Floaters & noise ↔ Drag

Quantitative Summary

Scene	PSNR ↑		SSIM ↑		LPIPS ↓
Scene	PSC	CF	PSC	CF	PSC	CF
Calib Lab	25.46	20.02	0.914	0.705	0.127	0.411
Dino	23.20	16.41	0.861	0.568	0.129	0.396
Eleutheria	24.11	17.67	0.807	0.634	0.275	0.466
Olympic Arch	24.01	17.50	0.913	0.693	0.089	0.341

Benchmark Datasets

PSC applied to raw calibration imagery from two public benchmarks — demonstrating that existing datasets can be recalibrated without any new hardware to improve downstream VO accuracy.

EuRoC-MAV — ATE RMSE (m)

Sequence	Mode	Default CF	PSC
MH01	VIO	0.074	0.038
MH02	VIO	0.209	0.065
MH03	VIO	0.193	0.087
MH02	Mono VO	45.03	10.55
MH05	Mono VO	12.21	3.33

Oxford Spires — ATE RMSE (m)

Sequence	Default CF	PSC
Keble College 02	0.858	0.418
Keble College 04	0.398	0.273
Observatory Qtr 01	0.341	0.137
Observatory Qtr 02	0.361	0.159

Citation

If this work is useful to your research, please consider citing:

@article{bharadwaj2026beyondcheckerboard, title = {Beyond Checkerboards: Advantages of Photogrammetric Camera Calibration for Robust 3D Vision and Novel View Generation}, author = {Bharadwaj, Akshay and Lichti, Derek D.}, journal = {International Journal of Computer Vision}, year = {2026}, note = {Under Review} }