Post-Undistortion Uncertainty Analysis

• Post-undistortion uncertainty is defined as the residual ambiguity in recovering latent signals after inverting a lossy, many-to-one transformation.
• It arises in systems such as radiometric correction (recovering RAW colors from tone-mapped JPEGs) and LiDAR deskewing (correcting scan distortions) using anisotropic, state-dependent covariance models.
• Modeling this uncertainty through Bayesian inversion or first-order covariance propagation enhances downstream tasks like HDR fusion, photometric stereo, and precise odometry.

Searching arXiv for the cited works and closely related terminology to ground the article. arXiv search query: "Post-Undistortion Uncertainty tone-mapped color images LiDAR-inertial odometry dropout injection IO-CUE domain drift calibration" Post-undistortion uncertainty denotes the residual uncertainty that remains after an observation has been “corrected” by inverting a distortion or rendering process. In the literature represented here, the term has two explicit technical instantiations. In radiometric vision, it refers to uncertainty in recovering latent linear RAW color from tone-mapped JPEG values, where the forward camera pipeline is compressive, bounded, and quantized, so inverse radiometric calibration is inherently one-to-many (Chakrabarti et al., 2013). In LiDAR-inertial odometry, it refers to the residual positional uncertainty of a point after scan deskewing, where imperfect motion compensation under vibration leaves each undistorted point with a point-wise covariance that should be propagated into matching and filtering (Dong et al., 6 Jul 2025). A broader reading, suggested by later post-hoc uncertainty work, is that whenever a fixed preprocessing or restoration stage is inverted or applied before inference, uncertainty should be attached to the corrected output rather than treating the correction as exact.

1. Scope, terminology, and problem class

In its most concrete form, post-undistortion uncertainty arises when a nonlinear forward process destroys information and a downstream system nevertheless requires a corrected representation. The radiometric case begins with latent linear sensor measurements $x \in [0,1]^3$, called RAW color, which are mapped by a global rendering pipeline into output-referred JPEG colors $y \in \{0,\dots,255\}^3$. The LiDAR case begins with raw scan points acquired at different times during a scan, which are deskewed to a common time using estimated motion. In both settings, the corrected quantity is not fully determined by the observed quantity, because the forward process is lossy or only approximately inverted (Chakrabarti et al., 2013, Dong et al., 6 Jul 2025).

This uncertainty is distinct from ordinary sensor noise. In the radiometric setting, the central source of ambiguity is many-to-one compression caused by tone mapping, clipping, gamut mapping, and 8-bit quantization. In the LiDAR setting, the central source is residual deskewing error after undistortion, caused by rapid and non-smooth motion, limited IMU sampling frequency, and IMU noise during intense vibration. In both cases, the uncertainty is state-dependent rather than uniform.

Setting	Observed quantity	Corrected latent quantity	Uncertainty representation
Radiometric derendering	JPEG color $y$	RAW color $x$	$p(x\mid y)$ or $\mathcal N(\mu(y),\Sigma(y))$
LiDAR deskewing	Raw point $^{L}\mathbf{p}'_j$ with IMU data	Undistorted point $^{L}\mathbf{p}_j$	$^{L}\mathbf{\Sigma}_{p_j}\in\mathbb{R}^{3\times 3}$

A recurring misconception is that undistortion simply produces a better point estimate. The literature instead treats the corrected value itself as uncertain. Another misconception is that uncertainty can be summarized by a scalar reliability score. Both the radiometric and LiDAR formulations use full anisotropic covariance structure, because some directions in the corrected space remain informative even when others are poorly constrained.

2. Probabilistic and covariance-based formulations

A general formulation begins with an explicit forward model $y=\mathbb J(x)$, an error model for mismatch in $y \in \{0,\dots,255\}^3$0, and a prior $y \in \{0,\dots,255\}^3$1. In the radiometric formulation, the inverse uncertainty is defined by

$y \in \{0,\dots,255\}^3$2

with

$y \in \{0,\dots,255\}^3$3

This is a Bayesian inversion of the full forward rendering model rather than a Jacobian-only local propagation scheme. The paper further approximates the inverse by

$y \in \{0,\dots,255\}^3$4

where $y \in \{0,\dots,255\}^3$5 is also the MMSE estimate under the exact posterior (Chakrabarti et al., 2013).

The LiDAR formulation adopts a different route. After deskewing a raw point by

$y \in \{0,\dots,255\}^3$6

the residual point error is linearized, and the post-undistortion covariance is assembled as

$y \in \{0,\dots,255\}^3$7

Here $y \in \{0,\dots,255\}^3$8 propagates angular deskewing error, $y \in \{0,\dots,255\}^3$9 captures translational deskewing error, and $y$0 is intrinsic LiDAR measurement covariance. The rotational term is

$y$1

which makes uncertainty explicitly range-dependent and anisotropic (Dong et al., 6 Jul 2025).

These two formulations exemplify two ends of a methodological spectrum. One is a posterior over latent corrected variables induced by a lossy forward map; the other is a first-order propagated covariance of the corrected variable itself. This suggests that post-undistortion uncertainty is not a single model class but a family of inverse-uncertainty constructions tied to the structure of the forward distortion and to the feasibility of exact Bayesian inversion.

3. Radiometric post-undistortion after tone mapping

The radiometric treatment in "Modeling Radiometric Uncertainty for Vision with Tone-mapped Color Images" formalizes post-undistortion uncertainty for digital photographs as the uncertainty of recovering linear scene-referred color from tone-mapped display-referred color (Chakrabarti et al., 2013). The forward rendering model is

$y$2

where $y$3 define a $y$4 linear color transform, $y$5 is a per-channel polynomial nonlinearity with $y$6 in the implementation,

$y$7

$y$8 are cross-channel corrective terms modeled with Gaussian RBF SVR,

$y$9

$x$0 bounds values to $x$1, and $x$2 quantizes to 8-bit integers.

The key claim is that deterministic inverse tone curves are incomplete because the forward pipeline is lossy. Near saturation, gamut boundaries, and strongly compressed regions of the tone curve, many distinct RAW colors map to the same JPEG color. The inverse is therefore represented as a probability distribution over linear scene colors rather than a single derendered value. For experiments, the prior is “a uniform prior over all possible sensor measurements whose chromaticities lie in the convex hull of the training data.” Errors in the estimated rendering function $x$3 are treated as Gaussian noise with variance $x$4, where $x$5 is set to twice the in-training RMSE.

To fit $x$6, the method uses corresponding RAW-JPEG pairs $x$7. To cover the reachable color and radiance domain, the calibration procedure captures an X-Rite 140-patch chart under sixteen different illuminants and multiple exposures from underexposed to overexposed. The standard transform plus polynomial stage is estimated by weighted least squares with monotonicity constraints on $x$8, alternating between quadratic programming for the polynomial coefficients and gradient descent for the color transform. A second stage fits gamut-correction residuals with support-vector regression.

The paper emphasizes that uncertainty varies substantially across color space. Covariance ellipsoids differ in both size and orientation; high variance can be concentrated largely along one RAW-space direction; variance tends to grow near output-gamut boundaries and saturation; and thresholding away bright or saturated pixels is arbitrary because uncertainty varies continuously across JPEG space. The per-pixel Gaussian approximation is computed numerically over a RAW grid, with reported runtime of roughly $x$9 ms per JPEG observation in MATLAB.

The practical relevance is established in three downstream tasks. In HDR/image fusion, each exposure contributes a Gaussian observation, and the latent color is estimated by precision-weighted fusion. In photometric stereo, uncertainty enters as heteroscedastic measurement variance, improving mean angular normal error on a Canon EOS 40D figurine sequence from $p(x\mid y)$0 for a deterministic calibrated inverse to $p(x\mid y)$1, while a naive gamma $p(x\mid y)$2 inverse gives $p(x\mid y)$3. In deconvolution, replacing a uniform fidelity term with a Mahalanobis-weighted term improves PSNR on all 24 blurred images, with median improvement over the deterministic baseline of $p(x\mid y)$4 dB. Validation also shows large gains in mean empirical log-likelihood over deterministic inverse surrogates for both RAW-capable and JPEG-only cameras.

Several limitations are explicit. The model assumes a stationary global tone-mapping pipeline; it does not explicitly model demosaicing, local tone mapping, vignetting, or detailed sensor-noise structure in inversion; calibration is easier for RAW-capable cameras than for JPEG-only cameras; and the Gaussian approximation may be insufficient if $p(x\mid y)$5 is strongly non-Gaussian or multimodal. Even so, the paper establishes a canonical inverse-uncertainty paradigm: after a lossy radiometric correction, the output should be represented by $p(x\mid y)$6 or by a full posterior, not by a single calibrated RGB value.

4. Point-wise geometric post-undistortion uncertainty in LiDAR-inertial odometry

The geometric counterpart is developed in "Vibration-aware Lidar-Inertial Odometry based on Point-wise Post-Undistortion Uncertainty," where the corrected variable is a deskewed LiDAR point rather than a radiometrically linearized color (Dong et al., 6 Jul 2025). The paper considers high-speed ground robots on unstructured terrain, where intense high-frequency vibration makes scan undistortion difficult because of rapid and non-smooth state changes and because IMU measurements are noisy and limited in sampling frequency. The estimated state follows a Fast-LIO-style iterated Kalman filter,

$p(x\mid y)$7

with standard inertial propagation. The novelty lies in the measurement model rather than in state augmentation.

After deskewing, the true point is related to the estimated point by rotational error, translational error, and transformed measurement noise. Under a small-angle approximation, the rotational error is modeled as

$p(x\mid y)$8

which yields

$p(x\mid y)$9

This term is anisotropic and grows with the point lever arm, so far-range points are especially sensitive to angular jitter. Translational undistortion error is modeled as

$\mathcal N(\mu(y),\Sigma(y))$0

which is independent of point range. Intrinsic LiDAR noise is represented by

$\mathcal N(\mu(y),\Sigma(y))$1

with $\mathcal N(\mu(y),\Sigma(y))$2. The full deskewed-point covariance is the sum of these three terms.

The deskewing-error variances are not derived from exact continuous-time inertial covariance propagation. Instead, the method estimates scan-level vibration intensity by mean absolute deviation of LiDAR-frame angular velocity and linear velocity,

$\mathcal N(\mu(y),\Sigma(y))$3

and then sets point-specific standard deviations as

$\mathcal N(\mu(y),\Sigma(y))$4

with $\mathcal N(\mu(y),\Sigma(y))$5 for 100–200 Hz IMUs. This makes uncertainty depend on timestamp within the scan, point geometry, and scan-level vibration intensity. End-of-scan points and long-range points acquire larger ellipsoids.

The covariance enters the odometry pipeline twice. First, correspondence selection is modified: a point’s $\mathcal N(\mu(y),\Sigma(y))$6 Euclidean nearest neighbors are found, then reranked by Mahalanobis distance under the point covariance, and the best $\mathcal N(\mu(y),\Sigma(y))$7 are retained for plane fitting. Second, the covariance is projected into point-to-plane residual variance,

$\mathcal N(\mu(y),\Sigma(y))$8

so the iterated Kalman filter update uses uncertainty-aware residual weighting. This is geometrically significant: uncertainty tangent to the plane need not strongly increase $\mathcal N(\mu(y),\Sigma(y))$9, whereas uncertainty along the normal directly reduces measurement confidence.

Empirically, the method is evaluated on a 3DoF vibration platform, public datasets, and a wheeled robot on uneven natural terrain. On the vibration platform, average translation error is reduced from $^{L}\mathbf{p}'_j$0 cm to $^{L}\mathbf{p}'_j$1 cm and average rotation error from $^{L}\mathbf{p}'_j$2 to $^{L}\mathbf{p}'_j$3. On selected public sequences it achieves the best performance on most cases, including NCD-05, NCD-06, and M2DGR street08. On the authors’ recordings, mean APE is best on all four sequences, including $^{L}\mathbf{p}'_j$4 m on sharp-turn and $^{L}\mathbf{p}'_j$5 m on collision. Ablation shows that removing uncertainty modeling or uncertainty-guided matching degrades performance.

The method remains real-time in the reported implementation. For scans downsampled to about 2,500 points, point uncertainty calculation takes about $^{L}\mathbf{p}'_j$6 ms, uncertainty-guided matching about $^{L}\mathbf{p}'_j$7 ms total, and total processing time per scan about $^{L}\mathbf{p}'_j$8 ms on a 3.8 GHz AMD CPU mini-PC. The main caveats are equally explicit: the vibration-to-uncertainty mapping is heuristic, the derivation relies on small-angle and first-order approximations, the errors are modeled as zero-mean Gaussian, and full covariance-aware nearest-neighbor search is approximated by Euclidean candidate retrieval followed by Mahalanobis reranking.

5. Post-hoc uncertainty retrofits for fixed processing pipelines

A separate literature does not model undistortion physics directly but addresses a nearby deployment problem: how to attach uncertainty after a fixed processing or regression stage without retraining the original model. This is relevant when an undistortion or restoration module is already deployed and the practical requirement is uncertainty on the downstream prediction rather than on the correction itself.

"Dropout Injection at Test Time for Post Hoc Uncertainty Quantification in Neural Networks" studies the case where a network is first trained without dropout, after which dropout layers are added and activated only at inference (Ledda et al., 2023). For Monte Carlo sample $^{L}\mathbf{p}'_j$9, weights become $^{L}\mathbf{p}_j$0, repeated stochastic passes yield a predictive mean and a variance $^{L}\mathbf{p}_j$1, and the predictive distribution is treated as Gaussian. The central finding is that raw injected-dropout variance is usually mis-scaled and must be rescaled: $^{L}\mathbf{p}_j$2 with NLL-optimal scale

$^{L}\mathbf{p}_j$3

The paper further introduces Miscalibration Area (MA) and a relaxed scale $^{L}\mathbf{p}_j$4 to navigate the NLL–calibration trade-off. In dense prediction, the viable dropout rates are extremely small, $^{L}\mathbf{p}_j$5, because larger rates sharply degrade RMSE. The practical lesson is that post-hoc uncertainty for a frozen image-to-image or regression model is feasible, but the variance is not plug-and-play: it requires scale calibration and often additional calibration adjustment.

"Principled Input-Output-Conditioned Post-Hoc Uncertainty Estimation for Regression Networks" proposes IO-CUE, which learns a separate uncertainty model

$^{L}\mathbf{p}_j$6

for a frozen predictor $^{L}\mathbf{p}_j$7, trained by detached Gaussian NLL on a labeled probe dataset (Bramlage et al., 1 Jun 2025). In canonical notation, the post-hoc objective is

$^{L}\mathbf{p}_j$8

The theoretical claim is that, under the Gaussian model and when the frozen mean predictor coincides with the conditional mean, this sequential fitting recovers the canonical heteroscedastic Gaussian MLE. The framework insists on conditioning on both the original input $^{L}\mathbf{p}_j$9 and the frozen output $^{L}\mathbf{\Sigma}_{p_j}\in\mathbb{R}^{3\times 3}$0: input conditioning preserves aleatoric heteroscedasticity, while output conditioning can support a relative or “quasi-epistemic” signal tied to the geometry of structured outputs. The paper explicitly states that conditioning on $^{L}\mathbf{\Sigma}_{p_j}\in\mathbb{R}^{3\times 3}$1 alone cannot recover input-dependent aleatoric variance and “could at best learn a global scalar variance.”

For post-undistortion systems, these works support a precise but limited interpretation. They directly support uncertainty on a frozen downstream prediction after an undistortion stage. They do not directly provide uncertainty of the undistorted image or corrected point cloud itself unless supervision exists at that stage. This suggests a useful taxonomy: uncertainty after undistortion can mean either uncertainty of the corrected signal itself, as in radiometric derendering and LiDAR deskewing, or uncertainty of a downstream prediction conditioned on a frozen corrected signal, as in post-hoc neural retrofits.

6. Calibration under domain drift and deployment after correction

A final issue is that post-undistortion outputs often do not follow the same distribution as the data on which confidence or uncertainty calibrators were originally fitted. "Post-hoc Uncertainty Calibration for Domain Drift Scenarios" studies this failure mode for multiclass image classification and shows that standard post-hoc calibration methods, when fitted on clean in-domain validation data, yield highly over-confident predictions under domain shift (Tomani et al., 2020). The proposed remedy is not a new calibrator family but a new calibration protocol: before fitting the calibrator, perturb the validation set so that model accuracy decreases linearly in 10 steps from the in-domain validation accuracy to random-guess accuracy $^{L}\mathbf{\Sigma}_{p_j}\in\mathbb{R}^{3\times 3}$2. Each perturbation level is produced by additive Gaussian noise $^{L}\mathbf{\Sigma}_{p_j}\in\mathbb{R}^{3\times 3}$3, with $^{L}\mathbf{\Sigma}_{p_j}\in\mathbb{R}^{3\times 3}$4 chosen by Nelder–Mead so that the pretrained model’s accuracy matches the target level. The calibrator is then fitted on the union of these perturbed sets.

The empirical pattern is consistent across CIFAR-10, ImageNet, and ObjectNet. Standard post-hoc calibration works in-domain but fails under drift; perturbation-based calibration improves shifted calibration substantially; and the gain generalizes to unseen shift types. For CIFAR-10 VGG19, mean ECE drops from $^{L}\mathbf{\Sigma}_{p_j}\in\mathbb{R}^{3\times 3}$5 with TS to $^{L}\mathbf{\Sigma}_{p_j}\in\mathbb{R}^{3\times 3}$6 with TS-P and to $^{L}\mathbf{\Sigma}_{p_j}\in\mathbb{R}^{3\times 3}$7 with IRM-P. For ImageNet ResNet50, mean ECE drops from $^{L}\mathbf{\Sigma}_{p_j}\in\mathbb{R}^{3\times 3}$8 with TS to $^{L}\mathbf{\Sigma}_{p_j}\in\mathbb{R}^{3\times 3}$9 with TS-IR-P. The paper also reports that the “-P” variants reduce confidence on OOD cases.

For post-undistortion uncertainty, the transferable implication is direct: when a correction stage changes image statistics, calibration on pristine validation data is insufficient. A plausible implication is that the validation set used to fit confidence maps, uncertainty heads, or post-hoc calibrators should be passed through the deployed correction pipeline or through realistic simulations of its residual artifacts. The paper explicitly suggests an analogue for transformation-heavy pipelines: construct a calibration set by applying the transformation, parameter sweeps, or residual artifact models, then fit the calibrator on those transformed samples rather than on unprocessed data alone.

Across the literature, three boundary conditions recur. First, uncertainty after correction is generally input-dependent and anisotropic, not uniform. Second, calibration quality is separable from nominal likelihood or error metrics: NLL-optimal scaling need not produce low calibration error, and in-domain calibration need not survive domain drift. Third, the object of uncertainty must be specified. Some methods quantify uncertainty of the corrected signal itself; others quantify uncertainty of downstream predictions conditioned on a frozen correction stage. Treating these as interchangeable obscures both the mathematics and the operational scope of the corresponding methods.