Reflectance Consistency Index

• Reflectance Consistency Index is a set of normalized measures assessing how invariant reflectance estimates remain when conditions like illumination, exposure, and viewing geometry change.
• It is applied in diverse settings including quaternion Retinex, multi-view reconstruction, UAV multispectral imaging, self-supervised intrinsic decomposition, and LiDAR-guided methods.
• Variations in formulation, normalization strategies, and sensitivity to noise or registration errors highlight challenges in standardizing the RCI across different applications.

Reflectance Consistency Index (RCI) denotes a family of quantitative measures for assessing how stable a reflectance estimate remains when acquisition conditions change. In the cited literature, the relevant changes include illumination interpolation in Retinex decomposition, exposure and illumination-color variation in self-supervised intrinsic decomposition, cross-view observation in multi-view reconstruction, view-geometry variation in UAV multispectral imaging, spatial irradiance fluctuation during field deployment, and cross-modal alignment between LiDAR-derived reflectance and RGB appearance. Only one work in this set explicitly introduces an index under the name “Reflectance Consistency Index,” namely the quaternion Retinex formulation of QRetinex-Net (Agaian et al., 22 Jul 2025). In the remaining works, the original papers either use other consistency measures or admit candidate RCI definitions that are constructed directly from their reported statistics or loss functions (Mélou et al., 2017, Seo et al., 2021, Kinoshita et al., 2021, Xie et al., 2024, Qin et al., 9 Jun 2026, Chen et al., 13 Mar 2026).

1. Scope of the concept and main formulations

Across the literature, “reflectance consistency” refers to agreement of reflectance under transformations that should ideally affect illumination, geometry, or modality rather than the material signal itself. In intrinsic decomposition and Retinex settings, reflectance is the illumination-invariant component. In multi-view and UAV remote-sensing settings, consistency instead concerns whether the same surface patch yields stable reflectance across views or observation geometries. In LiDAR-guided reconstruction, reflectance is treated as a lighting-invariant material channel aligned with RGB boundaries (Seo et al., 2021, Kinoshita et al., 2021, Mélou et al., 2017, Chen et al., 13 Mar 2026).

The resulting indices are therefore not uniform in construction. Some are unit-interval stability scores, some are ratios of extremes, and some are percentage improvements over a baseline coefficient of variation. The principal forms appearing in the cited works are summarized below.

Setting	Consistency quantity	Interpretation
Quaternion Retinex (Agaian et al., 22 Jul 2025)	$$1 - \sup_{p \in \Omega}\sigma^2(R_\alpha(p)) /(1/4)$$	$1$ means perfectly invariant reflectance
UAV multi-angular multispectral imaging (Qin et al., 9 Jun 2026)	$$RCI_{\min\text{–}\max}(\lambda)=r_{\min}(\lambda)/r_{\max}(\lambda)$$	$1$ means perfect radiometric consistency
Variational multi-view reflectance estimation (Mélou et al., 2017)	Normalized cross-view reflectance disagreement	Higher means better cross-view agreement
Deep Retinex / intrinsic decomposition (Seo et al., 2021, Kinoshita et al., 2021)	$1$ minus averaged pairwise reflectance discrepancy	Higher means stronger invariance across exposure or color
UAV reflectance generation under variable weather (Xie et al., 2024)	$100\% \times \left(1-CV_M/CV_B\right)$	Percentage consistency improvement over a baseline
LiDAR-guided Gaussian splatting (Chen et al., 13 Mar 2026)	No named RCI; reflectance RMSE and joint consistency losses	Lower error and better boundary alignment imply stronger consistency

Taken together, these formulations suggest that RCI is not a single standardized metric. A plausible implication is that the index is best understood as a normalization strategy applied to a task-specific notion of reflectance invariance.

2. Illumination-invariance RCI in quaternion Retinex

The most explicit and formal use of the term appears in QRetinex-Net, which replaces the classical channel-wise Retinex factorization with a quaternion-valued decomposition. An RGB image is embedded pixel-wise as a purely imaginary quaternion,

$S(x,y)=0+R(x,y)i+G(x,y)j+B(x,y)k,$

and the scene is modeled as

$S(x,y)=QR(x,y)\otimes QI(x,y),$

where $\otimes$ denotes the Hamilton product, $QR$ is the reflectance-like component, and $1$0 captures illumination (Agaian et al., 22 Jul 2025).

The RCI is then defined through controlled illumination interpolation. Given a registered pair $1$1, intermediate images are generated as

$1$2

For each pixel $1$3, the sample variance across reflectance estimates is

$1$4

Because the scalar reflectance proxy is assumed normalized to $1$5, the maximal possible variance is $1$6. The manuscript prints a raw normalized worst-case variance, but its stated interpretation and reported tables are consistent with the operational form

$1$7

Under this definition, $1$8 corresponds to zero variance, and $1$9 corresponds to maximal instability (Agaian et al., 22 Jul 2025).

The choice of the spatial supremum is central. It makes the index a worst-case measure: even small regions in which reflectance breaks invariance depress the score. The paper states that this is intended “to capture the most critical breakdown in reflectance invariance.” It also distinguishes evaluation from training: the network is trained with a direct reflectance consistency term,

$$RCI_{\min\text{–}\max}(\lambda)=r_{\min}(\lambda)/r_{\max}(\lambda)$$0

while RCI is reserved for evaluation (Agaian et al., 22 Jul 2025).

Experimentally, the paper reports on LOLv1:

• RetinexNet: RCI $$RCI_{\min\text{–}\max}(\lambda)=r_{\min}(\lambda)/r_{\max}(\lambda)$$1
• KIND++: RCI $$RCI_{\min\text{–}\max}(\lambda)=r_{\min}(\lambda)/r_{\max}(\lambda)$$2
• U-RetinexNet: RCI $$RCI_{\min\text{–}\max}(\lambda)=r_{\min}(\lambda)/r_{\max}(\lambda)$$3
• Diff-Retinex: RCI $$RCI_{\min\text{–}\max}(\lambda)=r_{\min}(\lambda)/r_{\max}(\lambda)$$4
• QRetinex-Net: RCI $$RCI_{\min\text{–}\max}(\lambda)=r_{\min}(\lambda)/r_{\max}(\lambda)$$5

These values are reported alongside PSNR, SSIM, and LPIPS, and the paper uses them to show that reflectance invariance is not equivalent to reconstruction fidelity. That distinction is one of the defining features of the RCI concept in Retinex literature (Agaian et al., 22 Jul 2025).

3. View-dependent consistency: multi-view estimation and UAV observation geometry

In multi-view reflectance estimation, consistency is formulated as agreement of reflectance for corresponding projections of the same surface point. In the variational model of multi-view reflectance estimation, one reflectance map per view is estimated,

$$RCI_{\min\text{–}\max}(\lambda)=r_{\min}(\lambda)/r_{\max}(\lambda)$$6

and the relaxed consistency term is

$$RCI_{\min\text{–}\max}(\lambda)=r_{\min}(\lambda)/r_{\max}(\lambda)$$7

where $$RCI_{\min\text{–}\max}(\lambda)=r_{\min}(\lambda)/r_{\max}(\lambda)$$8 is the correspondence indicator and $$RCI_{\min\text{–}\max}(\lambda)=r_{\min}(\lambda)/r_{\max}(\lambda)$$9 is the Huber loss. The paper does not define a named RCI, but it gives a natural basis for one:

$1$0

with a corresponding Huber-based variant. In that formulation, $1$1 means perfect cross-view agreement of reflectance over all correspondences (Mélou et al., 2017).

A different but related use appears in UAV multispectral imaging with strong multi-angular sampling. There, the issue is not decomposition from shading, but directional radiometric variability induced by BRDF effects. The geometry-aware workflow refines camera intrinsics and extrinsics via structure-from-motion, annotates homogeneous regions on an orthomosaic, and reprojects those ROIs into multiple raw sub-images to recover multi-band reflectance and observation geometry for the same ground targets. The viewing direction is computed from

$1$2

with zenith and azimuth geometry expressed through $1$3, $1$4, and the relative azimuth angle

$1$5

Under fixed illumination during a short flight, the observed reflectance per band is treated empirically as

$1$6

with $1$7 approximately constant (Qin et al., 9 Jun 2026).

Because the paper’s empirical statistic is the band-wise ratio of maximum to minimum reflectance across observation geometry, the derived consistency index is

$1$8

The equivalent range-normalized form,

$1$9

is numerically identical. Red-edge and NIR bands (B7–B10) have $1$0, yielding $1$1. Visible bands (B1–B6) have $1$2–$1$3, yielding $1$4 (Qin et al., 9 Jun 2026).

This is a different notion of consistency from the variational multi-view case. In the variational setting, high consistency is a property of estimated reflectance after disentangling shading. In the UAV BRDF setting, lower consistency directly reveals geometry-driven directional anisotropy of observed reflectance. The former treats inconsistency as an estimation defect; the latter treats it as a physical property of observation geometry.

4. Self-supervised intrinsic decomposition under exposure and illumination-color change

In self-supervised Retinex and intrinsic decomposition networks, reflectance consistency is defined through invariance across pseudo-images of the same scene generated with varied exposure or illumination color. One formulation decomposes an RGB image as

$1$5

and factorizes shading as

$1$6

where $1$7 is reflectance, $1$8 is gray-scale shading, and $1$9 is a global illumination color. The network enforces three consistencies: reconstruction consistency, reflectance consistency under exposure change, and reflectance consistency under illumination-color change. The pairwise reflectance loss is

$100\% \times \left(1-CV_M/CV_B\right)$0

with weights $100\% \times \left(1-CV_M/CV_B\right)$1, $100\% \times \left(1-CV_M/CV_B\right)$2, $100\% \times \left(1-CV_M/CV_B\right)$3, $100\% \times \left(1-CV_M/CV_B\right)$4, $100\% \times \left(1-CV_M/CV_B\right)$5, $100\% \times \left(1-CV_M/CV_B\right)$6, and $100\% \times \left(1-CV_M/CV_B\right)$7 in the experiments. The paper does not name an RCI, but it proposes a test-time index

$100\% \times \left(1-CV_M/CV_B\right)$8

together with exposure-specific and color-specific variants and a gradient-domain form $100\% \times \left(1-CV_M/CV_B\right)$9 (Seo et al., 2021).

A closely related intrinsic decomposition network adopts the color-illuminant model

$S(x,y)=0+R(x,y)i+G(x,y)j+B(x,y)k,$0

and enforces reflectance consistency with

$S(x,y)=0+R(x,y)i+G(x,y)j+B(x,y)k,$1

with $S(x,y)=0+R(x,y)i+G(x,y)j+B(x,y)k,$2, $S(x,y)=0+R(x,y)i+G(x,y)j+B(x,y)k,$3, $S(x,y)=0+R(x,y)i+G(x,y)j+B(x,y)k,$4, $S(x,y)=0+R(x,y)i+G(x,y)j+B(x,y)k,$5, and $S(x,y)=0+R(x,y)i+G(x,y)j+B(x,y)k,$6. Here again, the paper does not define a named RCI, but it gives several compatible forms. The one most directly aligned with the training loss is

$S(x,y)=0+R(x,y)i+G(x,y)j+B(x,y)k,$7

It also provides a variance-based form and an SSIM-based form (Kinoshita et al., 2021).

The experimental roles of these consistency mechanisms differ slightly. In the Deep Retinex network, reflectance consistency is linked to illumination-color estimation and white-balance adjustment; on 100 color-transferred images, the reported MSE improves from $S(x,y)=0+R(x,y)i+G(x,y)j+B(x,y)k,$8 to $S(x,y)=0+R(x,y)i+G(x,y)j+B(x,y)k,$9, and $S(x,y)=QR(x,y)\otimes QI(x,y),$0 improves from $S(x,y)=QR(x,y)\otimes QI(x,y),$1 to $S(x,y)=QR(x,y)\otimes QI(x,y),$2 (Seo et al., 2021). In the color-illuminant intrinsic decomposition network, reflectance consistency is evaluated through PSNR, MSE, and DSSIM between reflectance maps estimated from the same scene under different illuminants; on 45 simulated sets, the proposed method reports reflectance PSNR values in the 30–35 dB range, much lower MSE, and DSSIM around $S(x,y)=QR(x,y)\otimes QI(x,y),$3–$S(x,y)=QR(x,y)\otimes QI(x,y),$4 relative to the cited baselines (Kinoshita et al., 2021).

5. Spatial consistency in UAV reflectance generation under variable weather

A distinct use of reflectance consistency appears in UAV reflectance generation with an onboard downwelling spectrometer. The paper does not define a metric called RCI, but it explicitly quantifies consistency with the coefficient of variation,

$S(x,y)=QR(x,y)\otimes QI(x,y),$5

computed within a homogeneous vegetation area. In the cloudy-field UAV test, the vertical means across the short image dimension are first computed and then the CV of those means is measured along the plot direction. Reflectance is expressed in percent, and the CV is unitless but reported as a percentage; lower CV indicates higher spatial consistency (Xie et al., 2024).

The paper’s consistency improvement can be formalized as a baseline-relative index,

$S(x,y)=QR(x,y)\otimes QI(x,y),$6

and, when averaged across bands,

$S(x,y)=QR(x,y)\otimes QI(x,y),$7

This directly matches the paper’s interpretation of “improvement in reflectance consistency” (Xie et al., 2024).

The radiometric pipeline underlying this consistency measure is unusually explicit. A sinusoidal model corrects the downwelling spectrometer for solar-position variability,

$S(x,y)=QR(x,y)\otimes QI(x,y),$8

and the corrected DS signal is

$S(x,y)=QR(x,y)\otimes QI(x,y),$9

Band selection is then performed by PCA on the reference solar spectrum, and the adopted 4-Band MLR uses four 30 nm windows centered at $\otimes$0, $\otimes$1, $\otimes$2, and $\otimes$3 nm. For each MSI band $\otimes$4, the regression is

$\otimes$5

where $\otimes$6 are integrated ratio features constructed from MSI DN and corrected DS DN (Xie et al., 2024).

The central reported consistency numbers are:

• ELM baseline CV of the vertical mean reflectance along plot direction: $\otimes$7
• 4-Band MLR CV: $\otimes$8
• improvement:

$\otimes$9

The same method improves vegetation-index consistency in the same homogeneous rice plot. For

$QR$0

the reported CV decreases are $QR$1 for NDVI and $QR$2 for DVI, corresponding to improvements of $QR$3 and $QR$4, respectively. The paper also reports reflectance RMSE $QR$5 in a ground-based diurnal validation and $QR$6 in a UAV campaign conducted at various times throughout a sunny day (Xie et al., 2024).

6. Cross-modal reflectance consistency in LiDAR-guided Gaussian splatting

In LiDAR-guided Gaussian splatting for self-driving scenes, the paper explicitly states that it does not define or report a metric named “Reflectance Consistency Index.” Instead, reflectance consistency is enforced through calibration, supervision, and cross-modal alignment (Chen et al., 13 Mar 2026).

The starting point is a LiDAR intensity-to-reflectance calibration,

$QR$7

where $QR$8 is the LiDAR return intensity, $QR$9 is reflectance, $1$00 is range, and $1$01 is the incidence angle. The calibrated reflectance is projected into the image plane to form a sparse reflectance image $1$02. A reflectance-gradient image is also defined:

$1$03

These quantities supervise a per-Gaussian reflectance attribute $1$04, rendered by alpha blending as

$1$05

The LiDAR loss is

$1$06

with

$1$07

and weights $1$08 and $1$09 (Chen et al., 13 Mar 2026).

Reflectance consistency is further enforced at boundaries by a joint loss between reflectance and grayscale RGB gradients. After Gaussian smoothing,

$1$10

and

$1$11

with $1$12 and $1$13 (Chen et al., 13 Mar 2026).

Because no named RCI is given, the paper proposes only derived possibilities. One is a LiDAR-only aggregate of reflectance and reflectance-gradient agreement, and another is a boundary-consistency index based on average gradient cosine minus a weighted normalized magnitude gap. The paper’s own quantitative evidence is an ablation on Joint Loss: reflectance RMSE decreases from $1$14 without Joint Loss to $1$15 with Joint Loss, while PSNR improves from $1$16 to $1$17, SSIM from $1$18 to $1$19, and LPIPS from $1$20 to $1$21 (Chen et al., 13 Mar 2026).

The conceptual significance of this formulation is that reflectance consistency is neither purely geometric nor purely photometric. It is cross-modal: LiDAR-derived reflectance is treated as a lighting-invariant material prior, and RGB boundaries are used as an alignment constraint.

7. Interpretation, normalization, and limitations

The cited literature uses several normalization strategies. In QRetinex-Net, the normalizer is the maximal possible variance of a scalar reflectance in $1$22, namely $1$23, and the aggregation is a spatial supremum (Agaian et al., 22 Jul 2025). In multi-view variational estimation, normalization is by a global dynamic range or a variance-based constant $1$24, producing a unit-interval agreement score over geometry-derived correspondences (Mélou et al., 2017). In UAV multi-angular analysis, normalization is implicit in the ratio $1$25, so the index is bounded by $1$26 without distributional assumptions (Qin et al., 9 Jun 2026). In self-supervised intrinsic decomposition, normalization appears through pairwise distances, reflectance mean or luminance constraints, or SSIM-style similarity (Seo et al., 2021, Kinoshita et al., 2021). In field UAV reflectance generation, the normalization is relative to a baseline CV, so the resulting percentage expresses improvement rather than absolute invariance (Xie et al., 2024).

The interpretation of high and low values also differs. In the quaternion Retinex, multi-view variational, deep Retinex, and intrinsic-decomposition formulations, higher RCI means that reflectance is better separated from illumination and therefore more invariant. In the UAV BRDF case, a lower $1$27 indicates stronger directional anisotropy in observed reflectance, which is physically meaningful rather than algorithmically undesirable. In the UAV weather-calibration case, a higher baseline-relative RCI means stronger spatial stabilization under fluctuating irradiance. In the LiDAR-guided reconstruction case, the absence of a named scalar index leaves RMSE and cross-modal losses as the operative measures (Agaian et al., 22 Jul 2025, Qin et al., 9 Jun 2026, Xie et al., 2024, Chen et al., 13 Mar 2026).

The limitations reported in the literature are correspondingly heterogeneous. Worst-case aggregation by supremum is stringent and can be depressed by small misregistered regions or illumination-dependent noise in QRetinex-Net (Agaian et al., 22 Jul 2025). Fixed-illumination multi-view reflectance estimation is under-constrained and requires regularization; geometry errors affect normals, correspondences, and consistency (Mélou et al., 2017). UAV multi-angular reflectance depends on accurate pose, ground-plane normal, radiometric calibration, and illumination stability, and visible-band ratio metrics are sensitive when absolute reflectance is low (Qin et al., 9 Jun 2026). Self-supervised intrinsic models assume raw-domain linearity, diagonal chromatic adaptation, or a single global illumination color; mixed illuminants and strong non-Lambertian effects degrade reflectance consistency (Seo et al., 2021, Kinoshita et al., 2021). UAV reflectance generation with a downwelling spectrometer is sensitive to DS attitude, cosine-corrector non-idealities, solar altitude, cloud-driven spatial mismatch, spectral bandwidth choice, and BRDF (Xie et al., 2024). LiDAR-guided Gaussian splatting relies on incidence-angle correction, valid LiDAR-to-camera projection, and stable edge alignment; the paper does not model beam divergence, per-ring gain, or AGC/exposure explicitly (Chen et al., 13 Mar 2026).

Taken together, these works indicate that RCI is best understood as a problem-dependent formalization of a common physical and algorithmic desideratum: reflectance should remain stable when the changing factor is illumination, exposure, viewpoint, or sensing modality rather than the underlying material itself. The precise form of the index depends on which source of variation is treated as nuisance and which reflectance representation—scalar, RGB, quaternion, per-view map, per-band measurement, or per-Gaussian attribute—is being evaluated.