Pith. sign in

REVIEW 4 major objections 6 minor 22 references

Objective, Absolute and Hue-aware Metrics for Intrinsic Image Decomposition on Real-World Scenes: A Proof of Concept

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper proposes a way to measure absolute albedo in real scenes by combining hyperspectral images with LiDAR intensity, replacing subjective human annotations with a physically derived ground truth.

desk verdict A promising proof of concept for objective, hue-aware IID evaluation using hyperspectral + LiDAR albedo, but the absolute-scale claim rests on unreported calibration constants. read the letter →

arxiv 2505.19500 v1 pith:EIPJI4W5 submitted 2025-05-26 cs.CV

Shogo Sato , Masaru Tsuchida , Mariko Yamaguchi , Takuhiro Kaneko , Kazuhiko Murasaki , Taiga Yoshida , Ryuichi Tanida This is my paper · ORCID
classification cs.CV
keywords intrinsicimagedecompositionalbedohyperspectralimagingLiDARintensityradiometriccalibrationCIEDE2000spectralsimilarityhue-awareevaluation
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper proposes replacing human-judged relative-reflectance annotations (WHDR) with a directly measured albedo for evaluating intrinsic image decomposition on real scenes. The albedo is computed by combining hyperspectral images with LiDAR return intensity, exploiting the fact that the ratio of hyperspectral intensities at two wavelengths cancels the unknown surface geometry. On a reference color board with known albedo, the method achieves a CIEDE 2000 error of 6.75 and a luminance correlation of 0.981 with ground truth, suggesting that objective, absolute, and hue-aware IID evaluation is feasible outside synthetic datasets. An optional spectral-similarity step densifies the sparse LiDAR-based albedo for applications that need dense maps.

What carries the argument

The carrying object is the wavelength-ratio identity of Eqs. 3–4, which divides the hyperspectral intensity at wavelength $\lambda$ by the intensity at the LiDAR wavelength so that the geometry-dependent factor $m(n,l)$ cancels. The remaining factor is the ratio of incident-light spectra times the ratio of reflectances; with the illumination spectra calibrated from a white reference and the LiDAR reflectance $\rho(\lambda_{\mathrm{LiDAR}})$ obtained by inverting the LiDAR intensity model, this gives an absolute per-pixel reflectance spectrum. The supporting machinery is a non-parametric densification step: unknown pixels query a dictionary of known-albedo spectra using a hybrid Euclidean-plus-cosine distance and receive the average albedo of the three most similar dictionary entries.

What would settle it

Take a scene containing several flat color patches with known reflectance at different depths and incidence angles, illuminate it with a spatially varying spectrum, and compare the pipeline's computed albedo against a spectrometer measurement; if the CIEDE 2000 error grows past the 7.0 acceptance threshold as range, angle, or shadowing increases, the absolute-albedo claim is falsified.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is that absolute surface reflectance—and therefore ground-truth albedo—can be recovered for real scenes without manual annotation by fusing two active or calibrated measurements. The key operation is Eq. 4: for each pixel, the spectral reflectance $\rho(\lambda)$ equals $\frac{e(\lambda_{\mathrm{LiDAR}})}{e(\lambda)}\frac{I(\lambda)}{I(\lambda_{\mathrm{LiDAR}})}\rho(\lambda_{\mathrm{LiDAR}})$, where the ratio of hyperspectral pixel intensities cancels the common geometric shading factor $m(n,l)$, and $\rho(\lambda_{\mathrm{LiDAR}})$ is obtained from the LiDAR intensity equation. The recovered reflectance spectrum is converted to RGB via color matching functions, yielding an albedo map that removes cast shadows and can be compared against ground truth in absolute color units. The paper's experimental evidence is a color-board experiment in which the computed albedo outperforms the RGB image and existing IID baselines, with CIEDE 2000 error 6.75 (below the 7.0 threshold corresponding to 50% acceptance) and luminance correlation 0.981.

Load-bearing premise

The load-bearing premise is that the LiDAR intensity equation can be inverted to give an accurate absolute reflectance at the LiDAR wavelength, which requires known instrument constants, radiometric calibration, and reliable surface-normal estimates from sparse depth; a bias in this step flows through the ratio into every computed albedo and hue value.

Editorial extensions

If this is right

  • Real-world IID quality can be scored in absolute color units against a measured albedo, removing the subjectivity of WHDR annotations.
  • Hue is assessed quantitatively because the full reflectance spectrum is recovered, not just luminance.
  • Cast shadows do not corrupt the evaluation, since the geometry term is canceled by the wavelength ratio and LiDAR is an active sensor.
  • The sparse LiDAR-based albedo can be densified by spectral similarity, producing dense albedo maps suitable for training or full-image evaluation.
  • With denser LiDAR and multi-region illumination calibration, the same concept is intended to extend from the laboratory board to more complex real-world scenes.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the calibration assumptions hold outdoors, the same pipeline could turn any LiDAR-equipped platform into an albedo ground-truth collector, enabling fully supervised training of IID models on real imagery.
  • A hue-aware metric of this kind would expose color-shift errors that luminance-only or relative metrics miss, such as a method that gets brightness right but tints shadows blue.
  • One testable extension is to use the computed absolute albedo as a direct training signal, replacing pairwise human annotations with dense physical reflectance targets.
  • The densification step's reliance on spectral similarity suggests performance will degrade in scenes with many distinct materials but sparse LiDAR coverage; measuring that degradation is a natural next experiment.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes a quantitative evaluation protocol for intrinsic image decomposition on real scenes, replacing WHDR human annotations with physically measured albedo. The albedo at the LiDAR wavelength is obtained from LiDAR intensity via Eq. (1), hyperspectral ratios (Eqs. (3)-(4)) extrapolate it to the full spectrum, and the resulting reflectance is converted to RGB for CIEDE/CIE76 color-difference and luminance-correlation metrics. In a laboratory proof of concept on one Calibrite color board with cast shadows, the method reports CIEDE 2000 = 6.75, CIE76 = 13.5, and luminance correlation 0.981, along with an optional spectral-similarity densification step.

Significance. If the absolute calibration is trustworthy, the protocol would be a valuable advance: an objective, absolute, and hue-aware benchmark for real-world IID that does not rely on human relative judgments. The paper also offers a physically grounded densification procedure and compares against WHDR and several IID baselines. However, the current evidence is a single board, a single illumination condition, no error bars, and an undocumented LiDAR radiometric calibration chain; the significance is conditional on those issues being resolved. Credit is due for the internally consistent derivation of Eqs. (2)-(4), the honest limitation section, and the clear comparison with existing IID evaluation practice.

major comments (4)
  1. [Sec. 2.1, Eq. (1)] The paper never reports values or calibration procedures for the receiver aperture diameter D_r, system transmission factor eta_sys, or atmospheric attenuation eta_atm, nor does it state whether the Calibrite color board was used to determine these constants. If the board was used for calibration, Table 1 evaluates the method on the calibration target and the 6.75 CIEDE score is circular; if it was not, an unknown scale factor in rho(lambda_LiDAR) propagates linearly through Eq. (4) into every absolute albedo spectrum and affects CIEDE lightness. The authors must specify the full radiometric calibration chain, report the relevant constants or their estimation procedure, and ideally validate on a held-out target not used for calibration.
  2. [Sec. 3.3, Table 1] The headline result is a single measurement on one 24-patch board; there are no repeated acquisitions, no error bars, and no patch-level error analysis. Given that the reported CIEDE 2000 value of 6.75 is close to the cited acceptance threshold of about 7.0, the claim that the method 'comfortably exceeds' the threshold requires confidence intervals or multiple independent boards before it is fully supported.
  3. [Sec. 3.4, Fig. 7] The albedo densification method is one of the three stated contributions, but its evaluation is entirely visual. The claims of 'edge-preserved albedo with minimal artifacts' are unsupported without a quantitative metric, such as comparison against the measured sparse albedo on held-out points, or an error measure over patches with known ground-truth colors.
  4. [Sec. 3.5, item 2] The statement that real-world sunlight is 10-100 times brighter than the artificial sunlamp, making electronic noise negligible, is presented without measurement, citation, or numerical context. This quantitative generalization claim should be substantiated or softened, since it directly supports the extension of the method from the lab to outdoor scenes.
minor comments (6)
  1. [Fig. 2(b) caption] The caption contains a typo: 'dose not' should be 'does not'.
  2. [Eq. (5)] The weight parameter alpha is set empirically to 1.0, but no sensitivity analysis is provided; the effect of alpha on the dense albedo results in Table 1 should be stated.
  3. [Sec. 2.2] The procedure for estimating e(lambda) and e(lambda_LiDAR) from the white reference is described only briefly; please clarify whether the white reference is assumed perfectly Lambertian and how its known reflectance is accounted for.
  4. [Sec. 3.1] The paper does not state the Velodyne Alpha Prime laser wavelength or confirm that the SPECIM IQ has a valid spectral band at that wavelength; please describe the spectral resampling used in Eq. (4).
  5. [Table 1] The row labels beginning with 'No' are ambiguous; they appear to be enumeration numbers rather than method names and should be reformatted for clarity.
  6. [Sec. 3.3, Ref. [16]] The CIEDE 2000 acceptance threshold of about 7.0 is taken from a cockpit display study; the transfer of this threshold to intrinsic-image-decomposition evaluation should be justified.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the albedo is an independent physical measurement validated against an external color board.

full rationale

The paper's core derivation (Eqs. 1-4) is a physical measurement chain rather than a fit to the evaluation target. Eq. 4 computes rho(lambda) from measured hyperspectral intensities, a white-reference illumination calibration, and a LiDAR-derived rho(lambda_LiDAR); each term is independently measured or defined by the LiDAR radiometric model. The Calibrite color board is used as an external ground truth for evaluation, and the paper does not state that the LiDAR model constants Dr, eta_sys, eta_atm were calibrated on that board. Without such a statement, the reported CIEDE2000 error of 6.75 is evidence of accuracy, not a tautology. The only hand-set parameter (alpha = 1.0) enters the optional spectral-similarity densification, not the core albedo calculation, and densification is evaluated visually rather than presented as a predictive result. Self-citations [10,11] are prior LiDAR-IID works used as context or baselines, not as load-bearing premises of the derivation. The Sec. 3.5 limitation on surface-normal estimation and LiDAR intensity calibration is a correctness risk about uncalibrated constants, but it is not a circular step: the paper does not reduce its conclusion to an input. Therefore no specific circular reduction can be exhibited, and the derivation is self-contained against an external benchmark.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The central claim rests on radiometric calibration and Lambertian assumptions; the only hand-tuned free parameters belong to the optional densification step. No invented physical entities are introduced.

free parameters (2)
  • alpha (spectral similarity weight) = 1.0
    Set empirically in Eq. 5 for albedo densification; no sensitivity analysis or search reported.
  • neighborhood size k = 3
    Hand-selected averaging window in Eq. 6 for densification; no ablation.
assumptions (5)
  • domain assumption Lambertian surface model
    Eq. 2 assumes only diffuse reflectance and a geometry factor m(n,l) that cancels in the ratio; glossy or specular surfaces would bias the computed reflectance spectrum.
  • domain assumption LiDAR intensity model with known constants
    Eq. 1 requires receiver aperture, system transmission, atmospheric attenuation, and accurate incidence angle to recover absolute reflectance rho(lambda_LiDAR); calibration details are not fully provided.
  • domain assumption Incident light spectrum measured by whiteboard and treated as representative
    The ratio e(lambda_LiDAR)/e(lambda) in Eq. 4 is applied per pixel; spatially varying illumination would require multiple whiteboard measurements, as the paper notes in Sec. 3.5.
  • domain assumption Spatial registration of LiDAR points to hyperspectral pixels
    Albedo calculation assumes the sparse LiDAR returns are correctly aligned with hyperspectral image coordinates; alignment is mentioned but not detailed.
  • standard math Standard color matching functions
    Reflectance spectra are converted to RGB via XYZ color matching [12,13], a standard colorimetric operation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Objective, Absolute and Hue-aware Metrics for Intrinsic Image Decomposition on Real-World Scenes: A Proof of Concept." pith.science (2026). https://pith.science/paper/EIPJI4W5

@misc{pith2026250519500,
  author       = {Pith},
  title        = {Pith review of: Objective, Absolute and Hue-aware Metrics for Intrinsic Image Decomposition on Real-World Scenes: A Proof of Concept},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EIPJI4W5}},
  note         = {Machine review of arXiv:2505.19500}
}
read the original abstract

Intrinsic image decomposition (IID) is the task of separating an image into albedo and shade. In real-world scenes, it is difficult to quantitatively assess IID quality due to the unavailability of ground truth. The existing method provides the relative reflection intensities based on human-judged annotations. However, these annotations have challenges in subjectivity, relative evaluation, and hue non-assessment. To address these, we propose a concept of quantitative evaluation with a calculated albedo from a hyperspectral imaging and light detection and ranging (LiDAR) intensity. Additionally, we introduce an optional albedo densification approach based on spectral similarity. This paper conducted a concept verification in a laboratory environment, and suggested the feasibility of an objective, absolute, and hue-aware assessment. (This paper is accepted by IEEE ICIP 2025. )

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

22 extracted references · 21 canonical work pages

  1. [1]

    Objective, Absolute and Hue-aware Metrics for Intrinsic Image Decomposition on Real-World Scenes: A Proof of Concept

    INTRODUCTION Intrinsic image decomposition (IID) [1] is the task of sep- arating an image into illumination-invariant components (albedo) and illumination-variant components (shade). IID supports various high-level vision tasks such as semantic segmentation [2]. IID methods are typically divided into two categories: synthetic data decomposition [3, 4] and...

    work page Pith review arXiv 2025

  2. [2]

    Subsequently, we describe an albedo calculation process to evaluate IID quality

    METHODOLOGY This section firstly reviews LiDAR intensity and hyperspec- tral images. Subsequently, we describe an albedo calculation process to evaluate IID quality. Then, our albedo densifica- tion method based on spectral similarity is described. 2.1. LiDAR intensity LiDAR aims to measure depth from time lags between laser emission and detection. In add...

    work page

  3. [3]

    Experimental setting As a proof of concept, we conducted experiments using a hyperspectral camera, LiDAR, a color board, and an artificial illumination, as illustrated in Fig

    EXPERIMENTS 3.1. Experimental setting As a proof of concept, we conducted experiments using a hyperspectral camera, LiDAR, a color board, and an artificial illumination, as illustrated in Fig. 4. As the hyperspectral camera, we used the SPECIM IQ, which covers a broad wavelength range, including the LiDAR wavelength. For LiDAR, we employed the Velodyne Al...

    work page 2000

  4. [4]

    Our method objectively computes the ab- solute albedo values along with hue assessment

    CONCLUSION This paper proposed the concept of albedo calculation for as- sessing real-world IID using hyperspectral images and Li- DAR intensity. Our method objectively computes the ab- solute albedo values along with hue assessment. Achieving an error rate of approximately 6.75 (CIEDE 2000 standard), our proposal comfortably exceeds the acceptance thresh...

    work page 2000

  5. [5]

    Lightness and retinex theory,

    Edwin H. Land and John J. McCann, “Lightness and retinex theory,”Journal of the Optical Society of Amer- ica, vol. 61, no. 1, pp. 1–11, 1971

    work page 1971

  6. [6]

    A robust fruit image seg- mentation algorithm against varying illumination for vi- sion system of fruit harvesting robot,

    Chenglin Wang, Yunchao Tang, Xiangjun Zou, Weim- ing SiTu, and Wenxian Feng, “A robust fruit image seg- mentation algorithm against varying illumination for vi- sion system of fruit harvesting robot,”Optik, vol. 131, pp. 626–631, 2017

    work page 2017

  7. [7]

    CGIntrinsics: Better In- trinsic Image Decomposition through Physically-Based Rendering,

    Zhengqi Li and Noah Snavely, “CGIntrinsics: Better In- trinsic Image Decomposition through Physically-Based Rendering,” inECCV, 2018, pp. 371–387

    work page 2018

  8. [8]

    Uretinex-net: Retinex-based deep unfolding network for low-light im- age enhancement,

    Wenhui Wu, Jian Weng, Pingping Zhang, Xu Wang, Wenhan Yang, and Jianmin Jiang, “Uretinex-net: Retinex-based deep unfolding network for low-light im- age enhancement,” inCVPR, 2022, pp. 5901–5910

    work page 2022

Show all 22 references
  1. [9]

    Revisiting Deep Intrinsic Image De- compositions,

    Qingnan Fan, Jiaolong Yang, Gang Hua, Baoquan Chen, and David Wipf, “Revisiting Deep Intrinsic Image De- compositions,” inCVPR, 2018, pp. 8944–8952

    2018

  2. [10]

    NIID-Net: Adapting Surface Normal Knowledge for Intrinsic Image Decom- position in Indoor Scenes,

    Jundan Luo, Zhaoyang Huang, Yijin Li, Xiaowei Zhou, Guofeng Zhang, and Hujun Bao, “NIID-Net: Adapting Surface Normal Knowledge for Intrinsic Image Decom- position in Indoor Scenes,”IEEE TVCG, vol. 26, no. 12, pp. 3434–3445, 2020

    2020

  3. [11]

    ShapeNet: An Information-Rich 3D Model Repository,

    Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Xiao Jianx- iong, Li Yi, and Fisher Yu, “ShapeNet: An Information-Rich 3D Model Repository,”arXiv preprint arXiv:1512.03012, 2015

    2015 arXiv

  4. [12]

    Intrinsic Images in the Wild,

    Sean Bell, Kavita Bala, and Noah Snavely, “Intrinsic Images in the Wild,”ACM TOG, vol. 33, no. 4, pp. 1– 12, 2014

    2014

  5. [13]

    A review of LiDAR radio- metric processing: From ad hoc intensity correction to rigorous radiometric calibration,

    Alireza G Kashani, Michael J Olsen, Christopher E Par- rish, and Nicholas Wilson, “A review of LiDAR radio- metric processing: From ad hoc intensity correction to rigorous radiometric calibration,”Sensors, vol. 15, no. 11, pp. 28099–28128, 2015

    2015

  6. [14]

    Unsuper- vised Intrinsic Image Decomposition With LiDAR In- tensity,

    Shogo Sato, Yasuhiro Yao, Taiga Yoshida, Takuhiro Kaneko, Shingo Ando, and Jun Shimamura, “Unsuper- vised Intrinsic Image Decomposition With LiDAR In- tensity,” inCVPR, 2023, pp. 13466–13475

    2023

  7. [15]

    Unsupervised Single-Image Intrinsic Image Decompo- sition with LiDAR Intensity Enhanced Training,

    Shogo Sato, Takuhiro Kaneko, Kazuhiko Murasaki, Taiga Yoshida, Ryuichi Tanida, and Akisato Kimura, “Unsupervised Single-Image Intrinsic Image Decompo- sition with LiDAR Intensity Enhanced Training,” 2025, pp. 2368–2378

    2025

  8. [16]

    Creating RGB images from hyperspectral images using a color matching function,

    Magnus Magnusson, Jakob Sigurdsson, Sveinn Eirikur Armansson, Magnus O Ulfarsson, Hilda Deborah, and Johannes R Sveinsson, “Creating RGB images from hyperspectral images using a color matching function,” inIGARSS 2020-2020 IEEE International Geoscience and Remote Sensing Sympo...

    2020

  9. [17]

    40, John wiley & sons, 2000

    G ¨unther Wyszecki and Walter Stanley Stiles,Color sci- ence: concepts and methods, quantitative data and for- mulae, vol. 40, John wiley & sons, 2000

    2000

  10. [18]

    Ground truth dataset and baseline evaluations for intrinsic image algorithms,

    Roger Grosse, Micah K Johnson, Edward H Adelson, and William T Freeman, “Ground truth dataset and baseline evaluations for intrinsic image algorithms,” in ICCV, 2009, pp. 2335–2342

    2009

  11. [19]

    Learning Intrinsic Image Decomposition from Watching the World,

    Zhengqi Li and Noah Snavely, “Learning Intrinsic Image Decomposition from Watching the World,” in CVPR, 2018, pp. 9039–9048

    2018

  12. [20]

    Does External Illumination Affect Color Acceptability Threshold for a Mixed Display Technology Cockpit?,

    Pooshpanjan Roy Biswas, Dominique Dumortier, So- phie Jost, Herve Drezet, and Marie-Laure Avenel, “Does External Illumination Affect Color Acceptability Threshold for a Mixed Display Technology Cockpit?,” inLondon Imaging Meeting. Society for Imaging Sci- ence and Technology, ...

    2022

  13. [21]

    An image inpainting technique based on the fast marching method,

    Alexandru Telea, “An image inpainting technique based on the fast marching method,”Journal of graphics tools, vol. 9, no. 1, pp. 23–34, 2004

    2004

  14. [22]

    Deep Image Prior,

    Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky, “Deep Image Prior,” inCVPR, 2018, pp. 9446–9454

    2018

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.