REVIEW 3 major objections 5 minor 15 references
Fast Fourier Color Constancy and Grayness Index for ISPA Illumination Estimation Challenge
T0 review · 3 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash
Pith's one-line read Fourier method and gray-pixel take 3rd and 6th on Cube+ color constancy challenge.
desk verdict An honest, narrow challenge report whose headline FFCC result is weakened by an acknowledged mean-EXIF proxy; still worth a workshop referee. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
FFCC recasts color constancy as 2D localization in log-chroma space: log ratios $u=\log(I^g/I^r)$, $v=\log(I^g/I^b)$ map global illumination to a shift, and a convolutional filter on the UV histogram is applied as element-wise multiplication in Fourier space, giving fast inference. The Grayness Index instead applies a Laplacian-of-Gaussian filter to the log image; under locally constant illumination $\delta\log L^c(p)$ vanishes, so achromatic pixels satisfy $\delta\log I^r=\delta\log I^g=\delta\log I^b$, and the detected gray pixels vote for the illuminant. The paper's operative device for the FFCC entry is replacing the missing test-set EXIF with the mean training EXIF matrix, which lets a trained EXIF-dependent model run but degrades it relative to the no-EXIF variant.
What would settle it
Compare FFCC Model P on Cube+ training images using the mean EXIF matrix versus each image's own EXIF; a meaningful gap in median angular error confirms the substitution is the limiting factor, while no gap would falsify that explanation.
Extended reading notes
Core claim
On the undisclosed Cube+ test set, FFCC Model P achieves a median angular error of 1.64°, placing third, while the Grayness Index, used with no modifications and no learning, achieves 2.10°, placing sixth. The paper also shows the same methods on Gehler-Shi and NUS, where Cube+ proves the hardest benchmark, and attributes much of that difficulty to a large portion of local-patch images that break the gray-pixel assumption. The central finding is that these two off-the-shelf, non-deep methods transfer to a new benchmark with reasonable accuracy, and that missing EXIF metadata costs the Fourier method accuracy: the no-EXIF FFCC Model J finished ahead of the EXIF-dependent Model Q.
Load-bearing premise
The load-bearing assumption is that the mean EXIF matrix of training images adequately replaces the missing per-image EXIF for every test image, while the Grayness Index additionally assumes illumination is uniform within local neighborhoods.
Editorial extensions
If this is right
- FFCC can be executed on a benchmark whose test images omit EXIF by substituting the mean EXIF of the training set, at a measurable cost in accuracy.
- A learning-free statistical method can reach a median angular error within about 0.5° of a trained Fourier method on an undisclosed benchmark.
- Cube+ is a more demanding test bed than Gehler-Shi or NUS 8-camera, with local-patch images creating high-error outliers that inflate the mean.
- EXIF-based variants of FFCC rank below the no-EXIF variant when metadata is missing, confirming that biased metadata is worse than none.
Reading between the lines
- The mean-EXIF substitution probably sets a lower bound on what FFCC Model P could achieve with true per-image metadata; the true gap could be estimated by comparing mean-EXIF versus true-EXIF errors on Cube+'s training set.
- The same substitution trick could be applied to other sensor-specific pipelines at inference time, though its benefit will shrink when exposure settings vary widely across shots.
- GI's sensitivity to local patches suggests a natural training-free extension: detect close-up images and blend GI with a global gray-world estimate, which could reduce mean error without adding learned parameters.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. This short challenge report describes two submissions to the Illumination Estimation Challenge associated with ISPA 2019, evaluated on the Cube+ dataset. The first submission is Barron and Tsai's Fast Fourier Color Constancy (FFCC), specifically the variant the authors call Model P, which is reported to rank 3rd on the undisclosed Cube+ test set with a median angular error of 1.64. The second is the Grayness Index (GI) method from the authors' prior work, reported to rank 6th with a median angular error of 2.10. The paper explains that FFCC was trained using EXIF metadata and deep features, but because the released test images lacked EXIF, the authors substituted the mean EXIF matrix of the training images for every test image. The GI method is used as-is with no training. The paper also reports 3-fold cross-validation results on the Cube+ training set, the Gehler-Shi dataset, and the NUS 8-camera dataset, leading to the conclusion that Cube+ is the most challenging of the three datasets.
Significance. If the reported leaderboard positions are accurate, the paper demonstrates that two non-deep-learning methods, one requiring training and one learning-free, can achieve competitive rankings on a newly introduced color-constancy benchmark. The authors are candid about the EXIF substitution failure, and the use of publicly available code and open references supports reproducibility. The GI result is a useful data point for the learning-free approach. However, the manuscript contains no new methodological derivation, and its scientific contribution is limited to the challenge outcome plus a set of cross-dataset comparisons. The main value is as a workshop-style empirical report, provided the internal inconsistencies and the unquantified effect of the mean-EXIF proxy are addressed.
major comments (3)
- [Section III, Table I] The paper states that "we computed the mean exif matrix of all training images and used it for each testing image," but it does not quantify how much this constant-EXIF surrogate degrades FFCC Model P relative to using the correct per-image EXIF. The reported median angular error of 1.64 on the undisclosed Cube+ test data is therefore a property of a modified FFCC variant, not of standard FFCC Model P as validated on Gehler-Shi. Consequently, Section IV's claim that the comparison "validates that using a biased exif leads to a worse case" is not supported by the presented evidence, because no ablation with correct EXIF is reported. The authors should either add an ablation quantifying the EXIF-proxy degradation or reframe the 1.64 result explicitly as the outcome of the mean-EXIF-degraded variant rather than as standard FFCC Model P performance.
- [Section IV vs. Table I and Abstract] There is an internal inconsistency about which FFCC model was actually submitted. The Abstract and Table I identify the ranked-3rd submission as "FFCC Model P (our)," while Section IV states "Our FFCC-based Model Q suffers from the lack of exif for testing data, ranked after the FFCC Model J." If the submitted model was Model Q, then the table and abstract are mislabeled; if it was Model P, then the text is in error. Because the central ranking claim depends on exactly which model variant was evaluated, the authors must correct this discrepancy and use a single consistent model designation throughout the paper.
- [Section IV and Tables I-III] The conclusion that "the Cube+ dataset is the most challenging one" is based on a comparison that mixes evaluation protocols. The Cube+ undisclosed testing numbers are fixed leaderboard results on a held-out test set, while the Gehler-Shi (Table II) and NUS 8-camera (Table III) results are from 3-fold cross-validation on those datasets. In addition, the Cube+ training-data row for FFCC (median 0.84) is not directly comparable to the undisclosed-testing row. To support the "most challenging" claim, the authors should compare like with like, for example by using 3-fold cross-validation on the Cube+ training data versus the other two datasets, or explicitly label the cross-dataset comparison as informal and non-conclusive.
minor comments (5)
- [Section II, Eq. (2)] Equation (2) contains a typo: the expression for v(p) reads "log(I^g(p)/I^r(b))" but should presumably be "log(I^g(p)/I^b(p))." Please correct it.
- [Section II, text after Eq. (4)] The sentence "δ logI^r(p) = δ logI^g(p) = δ logI^b(p) indicates a perfect gray pixel at at the location p" contains a duplicated "at." Please fix the wording.
- [General] Several informal language issues appear, such as "warpped input images" instead of "warped," "This can be seemed as" instead of "This can be seen as," and "limitness" instead of "limitations." A careful proofreading pass is recommended.
- [Tables I-III] The 3-fold cross-validation results are reported without any measure of variance, such as standard deviation or confidence intervals for the median. Since the central comparison across datasets relies on these numbers, a variance estimate would strengthen the presentation.
- [Table I] The caption of Table I notes that the Cube+ training-data numbers are obtained by 3-fold cross-validation, but the table body does not visually distinguish the training rows from the undisclosed-test rows beyond placement. Adding a separator or explicit column headers such as "Undisclosed Testing Data" and "Cube+ Training Data" would improve clarity.
Circularity Check
No significant circularity: challenge rankings are external benchmark results, and the mean-EXIF substitution is a stated limitation, not a constructed prediction.
full rationale
This paper reports two existing color-constancy methods (Barron's FFCC and the authors' prior Grayness Index) applied to the ISPA challenge test set. The load-bearing results are external leaderboard rankings and angular-error statistics, produced by running public code on undisclosed test data; no parameter is fitted in this paper to those test labels and then re-derived as a prediction. The acknowledged mean-EXIF substitution ('we computed the mean exif matrix of all training images and used it for each testing image') is a stated limitation of the submitted FFCC configuration, not a hidden way of encoding the test-set answer into the method; similarly, the Model Q/P wording inconsistency affects claims about which variant was submitted, but it is not a circular construction. The only self-citation is to the authors' own Grayness Index paper [8], and it is not load-bearing in a circular sense: the challenge ranking for GI is a benchmark outcome generated by running the published code, not a conclusion derived from this paper's equations. Therefore no step reduces by construction to its own input, and the finding is no significant circularity.
Assumptions & free parameters
free parameters (2)
- FFCC hyperparameters (Model P) =
not reported
- Mean EXIF vector substitute =
average EXIF matrix over all Cube+ training images
assumptions (2)
- domain assumption Lambertian surfaces, narrow sensor response, and uniform global illumination (Eq. 1)
- domain assumption Illumination is constant over small local neighborhoods (Eq. 4)
Cite this review
Pith. "Pith review of Fast Fourier Color Constancy and Grayness Index for ISPA Illumination Estimation Challenge." pith.science (2026). https://pith.science/paper/4LXDPCHZ
@misc{pith2026190802076,
author = {Pith},
title = {Pith review of: Fast Fourier Color Constancy and Grayness Index for ISPA Illumination Estimation Challenge},
year = {2026},
howpublished = {\url{https://pith.science/paper/4LXDPCHZ}},
note = {Machine review of arXiv:1908.02076}
}
read the original abstract
We briefly introduce two submissions to the Illumination Estimation Challenge, in the Int'l Workshop on Color Vision, affiliated to the 11th Int'l Symposium on Image and Signal Processing and Analysis. The Fourier-transform-based submission is ranked 3rd, and the statistical Gray-pixel-based one ranked 6th.
Figures
Reference graph
Works this paper leans on
-
[1]
N. Bani ´c and S. Lon ˇcari´c. Unsupervised learning for color constancy. arXiv preprint arXiv:1712.00436 , 2017
arXiv 2017
-
[2]
J. T. Barron. Convolutional color constancy. In ICCV, 2015
work page 2015
-
[3]
J. T. Barron and Y .-T. Tsai. Fast fourier color constancy. In CVPR, 2017
work page 2017
-
[4]
S. Bianco and C. Cusano. Quasi-unsupervised color constancy. In CVPR, 2019
work page 2019
- [5]
-
[6]
Y . Hu, B. Wang, and S. Lin. Fully convolutional color constancy with confidence-weighted pooling. In CVPR, 2017
work page 2017
-
[7]
Y . Qian, K. Chen, J. K ¨am¨ar¨ainen, J. Nikkanen, and J. Matas. Recurrent color constancy. In ICCV, 2017
work page 2017
-
[8]
Y . Qian, J. Nikkanen, J. K ¨am¨ar¨ainen, and J. Matas. On finding gray pixels. In CVPR, 2019
work page 2019
Show all 15 references
-
[9]
Y . Qian, S. Pertuz, J. Nikkanen, J. K ¨am¨ar¨ainen, and J. Matas. Revisiting gray pixel for statistical illumination estimation. In International Conference on Computer Vision Theory and Applications , 2019
2019
-
[10]
Y . Qian, S. Yan, J. K ¨am¨ar¨ainen, and J. Matas. Flash lightens gray pixel. In ICIP, 2019
2019
-
[11]
Savchik, E
A. Savchik, E. I. Ershov, and S. M. Karpenko. Color cerberus. CoRR, abs/1907.06483, 2019
1907 arXiv
-
[12]
Shi and B
L. Shi and B. Funt. Re-processed version of the gehler color constancy dataset of 568 images. accessed from http:// www.cs.sfu.ca/ ∼colour/ data/, 2010
2010
-
[13]
W. Shi, C. C. Loy, and X. Tang. Deep specialized network for illumination estimation. In ECCV, 2016
2016
-
[14]
Yang, S.-B
K.-F. Yang, S.-B. Gao, and Y .-J. Li. Efficient illuminant estimation for color constancy using grey pixels. In CVPR, 2015
2015
-
[15]
B. Zhou, A. Khosla, A. Lapedriza, A. Torralba, and A. Oliva. Places: An image database for deep scene understanding. arXiv preprint arXiv:1610.02055, 2016
2016 arXiv
Reviewed August 14, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.