arxiv: 2604.10805 · v1 · submitted 2026-04-12 · 💻 cs.CV

Recognition: unknown

Analytical Modeling and Correction of Distance Error in Homography-Based Ground-Plane Mapping

Mateusz Szulc , Marcin Iwanowski

Authors on Pith no claims yet

Pith reviewed 2026-05-10 15:23 UTC · model grok-4.3

classification 💻 cs.CV

keywords homographydistance estimationground plane mappingerror modelingquadratic errormonocular cameracalibration correctiongradient descent

0 comments

The pith

Homography perturbations produce distance errors that grow quadratically with true ground distance from the camera.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper derives an explicit analytical link between small inaccuracies in manual homography initialization and the resulting systematic distance errors on the ground plane. It demonstrates that these errors increase approximately quadratically with the actual distance from the camera viewpoint. Two correction approaches are then tested: fitting a quadratic error function by regression and directly optimizing the homography parameters through coordinate-based gradient descent. Large-scale simulations involving over 19 million samples compare the methods, showing regression delivers higher peak accuracy when the model fits reliably while gradient descent offers better robustness to poor initial calibrations. The work concludes that refining geometric calibration often yields larger gains than adding model complexity in monocular monitoring applications.

Core claim

Small perturbations in a manually initialized homography that maps image coordinates to ground-plane positions create distance errors whose magnitude follows an approximately quadratic dependence on the true radial distance from the camera. This relationship is derived analytically and used to motivate two practical correction procedures: regression to estimate the quadratic error term and gradient-descent refinement of the homography itself. Simulation experiments confirm that the quadratic model captures the dominant distortion and that both correction strategies measurably reduce the resulting distance bias.

What carries the argument

The analytically derived quadratic mapping from homography perturbation parameters to distance error, which directly models the distortion as a function of ground distance.

If this is right

When the quadratic model can be fitted reliably, regression correction reduces distance error more effectively than uncorrected homography mapping.
Gradient-descent optimization of the homography parameters remains effective even when the initial manual calibration is substantially inaccurate.
In many practical systems, refining the geometric calibration step produces larger accuracy gains than increasing the complexity of downstream distance models.
The quadratic error relationship can be used to predict the magnitude of distance distortion at any given range before any correction is applied.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same quadratic dependence may appear in other homography-based tasks such as overhead traffic monitoring or robot navigation on flat surfaces.
A hybrid approach that first attempts regression and falls back to gradient descent could combine the accuracy and robustness benefits observed in the simulations.
The findings imply that calibration protocols should emphasize precise manual point placement rather than relying solely on post-hoc error modeling.

Load-bearing premise

The ground surface is a perfect plane and the dominant source of systematic distance error is the homography perturbation introduced by manual point selection.

What would settle it

A controlled experiment that applies known homography perturbations to a flat ground plane and measures distance errors that deviate significantly from the predicted quadratic curve would falsify the central model.

Figures

Figures reproduced from arXiv: 2604.10805 by Marcin Iwanowski, Mateusz Szulc.

**Figure 2.** Figure 2: This distribution defines the effective operating range of the system and serves as a representative test set for evaluating generalization beyond the calibration points [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

**Figure 3.** Figure 3: Example of distance-dependent calibration error, illustrating incon [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

**Figure 4.** Figure 4: Mean path improvement as a function of total distance sum and spread [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗

read the original abstract

Accurate distance estimation from monocular cameras is essential for intelligent monitoring systems. In many deployments, image coordinates are mapped to ground positions using planar homographies initialized by manual selection of corresponding regions. Small inaccuracies in this initialization propagate into systematic distance distortions. This paper derives an explicit relationship between homography perturbations and the resulting distance error, showing that the error grows approximately quadratically with the true distance from the camera. Based on this model, two simple correction strategies are evaluated: regression-based estimation of the quadratic error function and direct optimization of the homography via coordinate-based gradient descent. A large-scale simulation study with more than 19 million test samples demonstrates that regression achieves higher peak accuracy when the model is reliably fitted, whereas gradient descent provides greater robustness against poor initial calibration. This suggests that improving geometric calibration may yield greater performance gains than increasing model complexity in many practical systems.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Homography errors produce quadratic distance distortions; regression correction beats gradient descent in their 19M-sample simulation but everything stays simulated.

read the letter

The main thing to know is that this paper works out an analytical link between small homography perturbations and ground-plane distance error, showing the error grows roughly as the square of the true distance. They then compare two fixes: fitting a quadratic error function by regression versus tweaking the homography directly with gradient descent on point coordinates. The simulation says regression gives higher accuracy when the fit is reliable, while gradient descent holds up better against bad initializations.

Referee Report

2 major / 2 minor

Summary. The manuscript derives an explicit analytical relationship between small perturbations in a manually initialized homography and the induced distance errors on the ground plane, demonstrating approximate quadratic growth of error with distance. It then proposes and compares two correction approaches—regression fitting of the quadratic error function and gradient-descent optimization of the homography—using a large-scale simulation study comprising over 19 million samples to assess their accuracy and robustness.

Significance. If the perturbative derivation is valid within the range of typical manual homography errors, the work offers a lightweight, interpretable method to mitigate systematic biases in monocular distance estimation for surveillance and monitoring applications. The simulation provides quantitative evidence favoring regression when fitting is reliable and gradient descent for robustness, which could guide practical calibration improvements over model complexity increases.

major comments (2)

The small-perturbation expansion yielding the quadratic distance-error scaling is central to the claims, yet the manuscript provides no explicit bounds on perturbation size (e.g., in terms of pixel or angular error) for which higher-order terms can be neglected, nor does it compare these bounds to the magnitude of errors arising from manual point selection.
The 19-million-sample evaluation reports overall correction performance but omits stratified goodness-of-fit statistics (R² or residual analysis) for the quadratic model as a function of perturbation magnitude; this omission leaves the practical domain of the approximation unverified and weakens support for the quadratic claim under realistic conditions.

minor comments (2)

The abstract states 'regression achieves higher peak accuracy when the model is reliably fitted' but does not define the criterion for 'reliably fitted' or report the fraction of trials meeting it.
The distinction between the true homography H and the perturbed Ĥ should be introduced with explicit matrix equations early in the text to aid readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback and positive assessment of the work's potential impact. We address the two major comments point by point below. Both points identify valid opportunities to strengthen the validation of the perturbative model, and we will revise the manuscript accordingly.

read point-by-point responses

Referee: The small-perturbation expansion yielding the quadratic distance-error scaling is central to the claims, yet the manuscript provides no explicit bounds on perturbation size (e.g., in terms of pixel or angular error) for which higher-order terms can be neglected, nor does it compare these bounds to the magnitude of errors arising from manual point selection.

Authors: We agree that explicit bounds on perturbation size are needed to delineate the practical applicability of the quadratic approximation. In the revised manuscript we will add a dedicated analysis (new subsection and figure) that quantifies the range of homography perturbations—expressed both in pixel displacement at the image plane and in equivalent angular error—for which the first-order expansion matches the exact distance error to within 5 % relative error. This analysis will be performed by direct comparison of the perturbative formula against the full homography mapping over a dense grid of perturbation magnitudes. We will also relate these bounds to the typical 3–8 pixel localization uncertainty observed in manual point selection for ground-plane calibration, thereby grounding the theoretical limits in realistic operating conditions. revision: yes
Referee: The 19-million-sample evaluation reports overall correction performance but omits stratified goodness-of-fit statistics (R² or residual analysis) for the quadratic model as a function of perturbation magnitude; this omission leaves the practical domain of the approximation unverified and weakens support for the quadratic claim under realistic conditions.

Authors: We acknowledge the omission of stratified diagnostics. Although the simulation already sweeps perturbation magnitude systematically, the original manuscript reported only aggregate metrics. In the revision we will augment the evaluation section with binned goodness-of-fit results: R² values, mean absolute residuals, and residual histograms computed separately for perturbation intervals (e.g., 0–2 px, 2–5 px, 5–10 px, >10 px). These stratified statistics will be presented both in tabular form and as additional plots, directly verifying the domain over which the quadratic error model remains reliable. revision: yes

Circularity Check

0 steps flagged

Analytical derivation of quadratic distance error from homography perturbations is self-contained and independent of fitting

full rationale

The paper's core contribution is an explicit analytical relationship derived from homography properties via small-perturbation expansion, showing approximate quadratic growth of distance error with true distance. This is presented as a first-principles result from standard projective geometry mathematics rather than any fit to data or self-citation. The regression-based correction is a downstream application that estimates parameters of the already-derived quadratic form; it does not redefine or force the relationship itself. The 19-million-sample simulation evaluates correction performance but does not substitute for or circularly validate the derivation. No load-bearing step reduces to its own inputs by construction, and the derivation chain remains independent of the fitted correction strategies.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The work rests on standard assumptions of planar ground and homography validity plus the claim that manual initialization errors dominate; the regression strategy introduces fitted quadratic coefficients as free parameters.

free parameters (1)

quadratic error function coefficients
Fitted via regression in one correction strategy to match simulated distance errors.

axioms (2)

domain assumption The scene ground is a perfect plane and the homography represents the exact mapping when correctly initialized.
Required for any homography-based ground-plane mapping to be valid.
domain assumption Perturbations in homography initialization are small and produce systematic rather than random distance errors.
Underpins the derivation of the explicit quadratic relationship.

pith-pipeline@v0.9.0 · 5444 in / 1340 out tokens · 62464 ms · 2026-05-10T15:23:28.526854+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

22 extracted references · 21 canonical work pages

[1]

The Geometric Error for Homographies,

O. Chum et al., “The Geometric Error for Homographies,” Computer Vision and Image Understanding, vol. 97, no. 1,
[2]

doi:10.1016/j.cviu.2004.03.004

work page doi:10.1016/j.cviu.2004.03.004 2004
[3]

Virtual Worlds as Proxy for Multi- Object Tracking Analysis,

A. Gaidon et al., “Virtual Worlds as Proxy for Multi- Object Tracking Analysis,” inProc. IEEE CVPR, 2016. doi:10.1109/CVPR.2016.470

work page doi:10.1109/cvpr.2016.470 2016
[4]

Homography-based Correspondence in Weakly Calibrated Curved Surface Environment and its Error Analysis,

J. Liang and J. Su, “Homography-based Correspondence in Weakly Calibrated Curved Surface Environment and its Error Analysis,” inProc. IEEE International Conference on Robotics and Automation (ICRA), 2004, pp. 1203– 1208, doi:10.1109/ROBOT.2004.1307988

work page doi:10.1109/robot.2004.1307988 2004
[5]

A Multiview Approach to Tracking People in Crowded Scenes Using a Planar Ho- mography Constraint,

S. M. Khan and M. Shah, “A Multiview Approach to Tracking People in Crowded Scenes Using a Planar Ho- mography Constraint,” inComputer Vision – ECCV 2006, LNCS, vol. 3954, Springer, 2006, pp. 133–146, doi:10.1007/11744085_11

work page doi:10.1007/11744085_11 2006
[6]

Multi-camera Tracking and Segmentation of Occluded People on Ground Plane Using Search-Guided Particle Filtering,

K. Kim and L. S. Davis, “Multi-camera Tracking and Segmentation of Occluded People on Ground Plane Using Search-Guided Particle Filtering,” inComputer Vision – ECCV 2006, LNCS, vol. 3953, Springer, 2006, pp. 98– 109, doi:10.1007/11744078_8

work page doi:10.1007/11744078_8 2006
[7]

Tracking Multiple Occluding People by Localizing on Multiple Scene Planes,

S. M. Khan and M. Shah, “Tracking Multiple Occluding People by Localizing on Multiple Scene Planes,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 31, no. 3, pp. 505– 519, 2009, doi:10.1109/TPAMI.2008.102

work page doi:10.1109/tpami.2008.102 2009
[8]

Multiple camera peo- ple detection and tracking using support integration,

T. T. Santos and C. H. Morimoto, “Multiple camera peo- ple detection and tracking using support integration,”Pat- tern Recognition Letters, vol. 32, no. 1, pp. 47–55, 2011, doi:10.1016/j.patrec.2010.05.016

work page doi:10.1016/j.patrec.2010.05.016 2011
[9]

Multi camera image tracking,

J. Black and T. Ellis, “Multi camera image tracking,” Image and Vision Computing, vol. 24, no. 11, pp. 1256– 1267, 2006, doi:10.1016/j.imavis.2005.06.002

work page doi:10.1016/j.imavis.2005.06.002 2006
[10]

IEEE Transactions on Pattern Analysis and Machine Intelligence30(2), 267–282 (2008).https://doi.org/10.1109/TPAMI.2007.1174

F. Fleuret, J. Berclaz, R. Lengagne, and P. Fua, “Multicamera People Tracking with a Probabilis- tic Occupancy Map,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 30, no. 2, pp. 267–282, 2008, doi:10.1109/TPAMI.2007.1174

work page doi:10.1109/tpami.2007.1174 2008
[11]

Particle filtering with multiple and heterogeneous cameras,

R. Muñoz-Salinas, R. Medina-Carnicer, F. J. Madrid- Cuevas, and A. Carmona-Poyato, “Particle filtering with multiple and heterogeneous cameras,”Pattern Recognition, vol. 43, no. 7, pp. 2390–2405, 2010, doi:10.1016/j.patcog.2010.01.015

work page doi:10.1016/j.patcog.2010.01.015 2010
[12]

Calibra- tion and object correspondence in camera networks with widely separated overlapping views,

F. Yin, D. Makris, S. A. Velastin, and T. Ellis, “Calibra- tion and object correspondence in camera networks with widely separated overlapping views,”IET Computer Vi- sion, vol. 9, no. 3, pp. 354–367, 2015, doi:10.1049/iet- cvi.2013.0301

work page doi:10.1049/iet- 2015
[13]

Stereo inverse perspective mapping: theory and applications,

M. Bertozzi, A. Broggi, and A. Fascioli, “Stereo inverse perspective mapping: theory and applications,”Image and Vision Computing, vol. 16, no. 8, pp. 585–590, 1998, doi:10.1016/S0262-8856(97)00093-0

work page doi:10.1016/s0262-8856(97)00093-0 1998
[14]

Multimodal inverse perspective mapping,

M. Oliveira, V . M. F. Santos, and A. D. Sappa, “Multimodal inverse perspective mapping,”In- formation Fusion, vol. 24, pp. 108–121, 2015, doi:10.1016/j.inffus.2014.09.003

work page doi:10.1016/j.inffus.2014.09.003 2015
[15]

The Right (Angled) Perspective: Improving the Understand- ing of Road Scenes Using Boosted Inverse Perspective Mapping,

T. Bruls, H. Porav, L. Kunze, and P. Newman, “The Right (Angled) Perspective: Improving the Understand- ing of Road Scenes Using Boosted Inverse Perspective Mapping,” inProc. IEEE Intelligent V ehicles Symposium (IV), 2019, pp. 302–309, doi:10.1109/IVS.2019.8814056

work page doi:10.1109/ivs.2019.8814056 2019
[16]

An Inverse Per- spective Mapping Approach using Monocular Camera of Pepper Humanoid Robot to Determine the Posi- tion of Other Moving Robot in Plane,

M. H. Tanveer and A. Sgorbissa, “An Inverse Per- spective Mapping Approach using Monocular Camera of Pepper Humanoid Robot to Determine the Posi- tion of Other Moving Robot in Plane,” inProc. In- ternational Conference on Pattern Recognition Appli- cations and Methods (ICPRAM), 2018, pp. 219–225, doi:10.5220/0006930002190225

work page doi:10.5220/0006930002190225 2018
[17]

Camera Calibration from Video of a Walking Human,

F. Lv, T. Zhao, and R. Nevatia, “Camera Calibration from Video of a Walking Human,”IEEE Trans. Pattern Anal. Mach. Intell., vol. 28, no. 9, pp. 1513–1518, 2006, doi:10.1109/TPAMI.2006.178

work page doi:10.1109/tpami.2006.178 2006
[18]

Simultaneous surveillance camera calibration and foot-head homology estimation from human detections,

B. Micušík and T. Pajdla, “Simultaneous surveillance camera calibration and foot-head homology estimation from human detections,” inProc. IEEE CVPR, 2010, pp. 1562–1569. doi:10.1109/CVPR.2010.5539786

work page doi:10.1109/cvpr.2010.5539786 2010
[19]

Automatic Camera Calibration for Traffic Understanding,

M. Dubská, A. Herout, and J. Sochor, “Automatic Camera Calibration for Traffic Understanding,” inProc. BMVC, 2014, doi:10.5244/C.28.42

work page doi:10.5244/c.28.42 2014
[20]

AutoCalib: Automatic Traffic Cam- era Calibration at Scale,

R. Bhardwaj, G. K. Tummala, G. Ramalingam, R. Ram- jee, and P. Sinha, “AutoCalib: Automatic Traffic Cam- era Calibration at Scale,” inProc. ACM BuildSys, 2017, doi:10.1145/3137133.3137149

work page doi:10.1145/3137133.3137149 2017
[21]

CCTV-Calib: a toolbox to calibrate surveillance cam- eras around the globe,

F. Rameau, J. Choe, F. Pan, S. Lee, and I. S. Kweon, “CCTV-Calib: a toolbox to calibrate surveillance cam- eras around the globe,”Machine Vision and Applica- tions, vol. 34, article 125, 2023, doi:10.1007/s00138-023- 01476-1

work page doi:10.1007/s00138-023- 2023
[22]

One homog- raphy is all you need: IMM-based joint homography and multiple object state estimation,

P. J. Claasen and J. P. de Villiers, “One homog- raphy is all you need: IMM-based joint homography and multiple object state estimation,”Expert Sys- tems with Applications, vol. 302, article 130562, 2025, doi:10.1016/j.eswa.2025.130562. 7

work page doi:10.1016/j.eswa.2025.130562 2025