{"id":"c3653dd5-c17f-4310-a1da-f5f7ce27340e","arxiv_id":"2502.03494","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"IFFCC extends FFCC to multi-illuminant scenes by using an integral UV histogram and batched FFT convolution to compute local illumination estimates fast enough for real-time use.","lead":"A white-balance algorithm that estimates a smooth, region-by-region illumination map for scenes with multiple light sources, built by adding an integral-histogram accelerator to the Fast Fourier Color Constancy method. It runs in milliseconds on a CPU with a tiny parameter count, making real-time multi-illuminant auto white balance plausible on phones.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Single circular mean per window (Eq. 12) collapses multi-illuminant windows to one point, limiting IFFCC's accuracy on sharp boundaries; LSMI results show AID outperforming it.","rationale":"I align with the reader's conditional verdict. The integral UV histogram acceleration (Sec. 3.1) is a sound and useful contribution, and the runtime advantages on CPU are real and well-supported by Tables 3-5. However, the accuracy claim is stronger than the representation permits. The reader's weakest assumption—the single circular mean per window—is indeed the load-bearing limitation for the 'surpasses pixel-level networks' claim. The proposed concrete test would isolate whether the performance gap on mixed-illumination regions is caused by the per-window point estimate. If confirmed, the paper should qualify its claims to 'competitive accuracy at much lower cost' rather than 'on par with or surpassing' pixel-level networks. Since this is a claim-level correction rather than an invalidation of the algorithmic contribution, the condition to accept is to revise the accuracy claims and discuss the boundary-resolution limitation.","tokens_in":13321,"tokens_out":5845,"duration_ms":56105,"concrete_test":"On the LSMI test set, compute each pixel's GT illumination variance over its enclosing 128x128 window. Partition pixels into high-variance (window contains multiple illuminants) and low-variance sets. Report mean angular error for IFFCC and AID separately on each set. If IFFCC's error is comparable to AID on low-variance pixels but significantly worse on high-variance pixels, the single circular mean of Eq. 12 is the limiting factor; if the error gap is uniform, the bottleneck lies elsewhere.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that IFFCC matches or surpasses pixel-level networks rests on the assumption in Sec. 3.3 that a single circular mean (Eq. 12) of each window's filtered histogram is a sufficient local illuminant estimate. IFFCC uses 128x128 windows with 64 overlap, producing only a 3x3 grid of estimates on the 256x256 test images, then interpolates and guided-filters. Any window containing two distinct light sources is collapsed to one point; because the model is trained to predict the blended illumination within each window (Sec. 4.1), it does not learn to resolve the two illuminants. Guided filtering can sharpen edges but cannot introduce a hard boundary that was absent from the 3x3 grid. This is a representational ceiling, not a tuning issue: no increase in overlap or filter capacity within the current per-window single-point formulation recovers pixel-level precision at mixed-illuminant boundaries. The LSMI tables (Tables 3-4) are consistent with this: AID, a pixel-level slot-attention model, achieves lower mean angular error on every camera (e.g., 2.03 vs 2.48 on Canon Multi). Thus the abstract's 'on par with or surpasses' claim is not supported by the paper's own numbers, and the gap is explained by the single-mean collapse.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes IFFCC, a multi-illuminant extension of FFCC. It builds an integral UV histogram so that the log-chroma histogram of any rectangular window can be obtained by a few additions and subtractions, then applies FFT-based filtering to all window histograms in parallel. Per-window illuminants are recovered as circular means (Eq. 12), and the resulting coarse map is interpolated and guided-filtered. Experiments on the Shadow and LSMI datasets report angular errors, parameter counts, and CPU runtimes. The paper claims accuracy on par with or better than pixel-level neural networks, with over 400x fewer parameters and 20-100x less processing time.","tokens_in":13607,"tokens_out":11439,"duration_ms":98997,"significance":"The integral-histogram formulation is exact, and the reported CPU-only runtime (0.03 s on 256x256 images, 0.012M parameters) would be useful for embedded AWB if the accuracy claims were accurate. The Shadow dataset results are competitive, and the ablation study gives informative window-size trade-offs. However, the two headline claims (accuracy parity with pixel-level networks and 20-100x speedups) are not supported by the reported numbers: AID is consistently better on LSMI, and speedups over the listed network baselines range from about 9x to 33x, with FFCC-vs-IFFCC speedups of only 3-17x in Table 5. The contribution is still publishable after the claims are revised.","major_comments":[{"comment":"The claim that IFFCC achieves accuracy 'on par with or surpasses that of pixel-level neural networks' is contradicted by the paper's own experiments. On both LSMI subsets and for every camera, AID reports a lower mean angular error than IFFCC (e.g., Multi/Galaxy: IFFCC 2.48 vs AID 2.03; Mixed/Galaxy: IFFCC 1.98 vs AID 1.63), and the median is also better for AID in all six camera columns. The text should either add a comparison that actually supports the claim or replace it with a more modest statement such as 'competitive with, though not surpassing, leading pixel-level methods.'","section":"Abstract, Conclusion, Tables 3 and 4"},{"comment":"The '20-100x faster' speed claim is not supported by the reported runtimes. On LSMI, IFFCC's 0.03 s gives about 9x over LSMI-H (0.28 s), 16x over Pix2Pix (0.48 s), 17x over LSMI-U (0.51 s), and only vs AID (>1 s) is the speedup above 20x. In Table 5, the measured speedups over FFCC for the six window/overlap settings are about 3-17x, not 20-100x. Please qualify the speed claim by baseline and input size, or correct the numbers.","section":"Abstract, Conclusion, Tables 3-5"},{"comment":"Representing each window's filtered histogram by a single circular mean creates a structural ceiling for boundary accuracy. With the default window size 128 and overlap 64 on 256x256 test images, there are only 3x3 raw estimates; interpolation and guided filtering (Section 3.4) can smooth or align those estimates but cannot recover a hard illumination boundary that is absent from that coarse grid. Training against the blended illumination inside each window makes the model predict an average rather than resolve multi-modal windows. This is the likely reason AID, a pixel-level method, consistently outperforms IFFCC on LSMI, and it should be discussed as a limitation.","section":"Section 3.3, Eq. (12); Section 4.1"},{"comment":"The equation sums over the window index k inside the inverse FFT, which would collapse all window histograms into a single output; Eq. (13) then refers to P^k, which is only defined if Eq. (10) is interpreted per window. This is inconsistent as written and should be clarified (e.g., by writing the operation as a batched FFT without a sum over k, or by making the summation index another quantity). The current text makes the parallel-prediction recipe ambiguous and difficult to reproduce.","section":"Section 3.3, Eq. (10)"}],"minor_comments":[{"comment":"The union symbol is not the correct operation for accumulating histogram bin counts; these equations should use bin-wise addition, as Eq. (8) does.","section":"Section 3.1, Eqs. (5) and (7)"},{"comment":"The method computes independent circular means of the row and column marginals rather than the Bivariate von Mises mean discussed earlier in the section; the text should state this explicitly and justify the independence assumption.","section":"Section 3.3, Eqs. (12)-(13)"},{"comment":"The training loss and optimizer hyperparameters are not specified; 'training runs for 64 iterations' is insufficient for reproduction.","section":"Section 4.1"},{"comment":"The guided-filter window radius and regularization parameter (epsilon) are not reported, although they affect the final illumination map.","section":"Section 3.4"},{"comment":"References [9] and [10] are duplicate citations of Buchsbaum (1980) with different titles; the list should be deduplicated.","section":"References"},{"comment":"The caption says the time is for 256x256 images on a CPU, but it is not clear whether the timed 0.03 s includes interpolation and guided filtering; please specify the exact timing scope.","section":"Tables 3 and 4"},{"comment":"The constants 217, 274, and 11(68+4B2) are unexplained; please provide a derivation or remove the formula if it is not used further.","section":"Equation (9)"}],"recommendation":"major_revision","confidential_remarks":"This is an incremental but potentially useful engineering contribution. The integral-histogram idea is sound and the Shadow results are good, but the LSMI numbers directly undercut the abstract and conclusion. I recommended major revision mainly to force a rewriting of the headline claims and a cleanup of the parallel-prediction equation; no concerns about novelty disclosure or citation practice beyond the reference duplicates."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe one thing to know: IFFCC is a real, working extension of FFCC to local illumination estimation, built on Porikli's integral histogram. The math is checkable and the speedup idea is sound. But the paper's headline claim — 'on par with or surpasses pixel-level networks' — is not supported by its own Tables 3 and 4, where AID beats it on every LSMI camera. The stress-test note is right about the single circular mean being a representational ceiling, and the paper basically admits it by training on blended per-window GT.\n\nWhat's new: combining integral UV histograms with FFCC-style Fourier convolution, and parallelizing the BVM mean over a grid of windows, giving a CPU-friendly local AWB with few parameters. That is a legitimate extension, not a new framework. The integral-histogram equations (Eq. 8, 10-13) look correct, and the ablation study is honest about window-size tradeoffs.\n\nSoft spots, in proportion. First, the empirical overclaim: on LSMI Multi, AID gets 2.03 mean on Galaxy vs IFFCC's 2.48; on Mixed, 1.63 vs 1.98. So the 'comparable or surpasses' statement in the abstract is simply wrong for the best pixel-level method. Second, the speedup claim in the abstract and conclusion says 20-100x faster than network-based approaches, but Table 5 only compares to FFCC, not to AID or LSMI-H; the FFCC speedup is roughly 9x for typical window sizes, not 20-100x. Third, the single-mean collapse: with 128x128 windows on 256x256 images, you get a 3x3 grid of estimates. Guided filtering can sharpen but cannot manufacture a light boundary absent from that grid. The paper's defense — training on blended GT — makes the model predict the average, which is exactly what limits boundary precision. This is a real limitation, though it's not fatal for the method's stated goal: smooth, real-time preview on embedded devices. There, 2-3 degree mean angular error may be acceptable.\n\nThe citation pattern is fine; the prior work (Porikli, Barron, LSMI, AID) is properly credited. No circularity to speak of. The writing is uneven and some numbers are sloppy, but the core idea is coherent.\n\nWho gets value: anyone working on practical AWB for phones or video, and people who want a fast baseline for local color constancy. It's worth a serious referee — the method is reproducible enough and the results are informative, but the abstract and conclusion need rewording to match the data.\n\nMy recommendation: send it to review, but require the authors to report AID as a baseline in the same tables, fix the speedup claim, and either relax the 'surpasses' language or add experiments with smaller windows to show the trade-off explicitly.","headline":"Solid integral-histogram extension of FFCC with honest math, but the paper overclaims accuracy and speed relative to its own LSMI tables.","tokens_in":14118,"tokens_out":1918,"would_cite":true,"duration_ms":16900,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a local extension of FFCC, built on an integral UV histogram and parallel FFT convolution, performs multi-illuminant white balance with accuracy matching pixel-level neural networks at over 400x fewer parameters and…","keywords":["integral histogram","fast Fourier color constancy","multi-illuminant auto white balance","color constancy","local illumination estimation","guided filtering","log-chromaticity","real-time image signal processing"],"falsifier":"On a synthetic two-illuminant image with a straight, sharp boundary and well-separated chromaticities, inspect each window's predicted illuminant against the two known ground-truth lights: if the per-window circular mean (Eq. 12) repeatedly falls between the two true chromaticities rather than near one of them, or if the reconstructed map's angular error on the boundary strip stays well above pixel-level networks, then the local-estimate assumption is the limiting factor.","tokens_in":13117,"feed_emoji":"🎨","tokens_out":6276,"duration_ms":58894,"temperature":0.7,"pith_summary":"This paper proposes a modification of the Fast Fourier Color Constancy (FFCC) algorithm so that it can handle scenes with several light sources at once. Instead of building a fresh histogram for every image region, IFFCC precomputes one integral UV histogram over the whole image in log-chromaticity space, then reads out any window's histogram with a few arithmetic operations. Illumination for all windows is estimated in parallel by fast Fourier convolution and circular-mean statistics, then interpolated and edge-preserved by guided filtering. The claim is that this reaches the accuracy of pixel-level neural networks on the LSMI and Shadow benchmarks with about 400x fewer parameters and 20-100x faster runtime. If correct, the result is a practical local auto white balance for real-time capture on CPUs and mobile devices.","feed_headline":"Local white balance matches neural nets at 400x lower cost","feed_subtitle":"Integral histograms plus parallel FFT put multi-illuminant AWB inside the 10ms real-time budget.","key_machinery":"The load-bearing object is the integral UV histogram: a recursive prefix-sum over log-chroma bins, written in Eq. 8, that turns any rectangular window's histogram into three vector additions and one subtraction. Around it, the method runs a parallelized fast Fourier convolution (Eq. 10), a softmax over the toroidal likelihood, and a parallel circular-mean step (Eq. 12) for each window's Bivariate von Mises estimate. This combination removes the repeated per-region histogram extraction that made direct FFCC extension slow and lets many local illuminant hypotheses be produced in a single inference pass.","core_discovery":"The central claim is that local multi-illuminant white balance does not need a pixel-level neural network: the same log-chroma histogram that FFCC uses globally can be made local and fast. The paper's integral UV histogram (Eq. 8) lets any rectangular region's histogram be computed from three previously scanned histograms, so a whole overlapping grid of windows is nearly free. Fourier-based convolution in the log-chroma torus and the Bivariate von Mises circular means are evaluated for all windows in one parallel pass, giving one illuminant estimate per window. After linear interpolation and guided filtering, those estimates become a spatially smooth but edge-preserving illumination map that can be directly applied to correct the image. On the Multi and Mixed splits of LSMI, IFFCC reports mean angular errors of 2.48/2.30/2.48 and 1.98/2.11/1.93 across Galaxy/Nikon/Sony cameras, in the same range as the pixel-level AID model, while using roughly 0.012M parameters and 0.03s on CPU.","pith_inferences":["The single-circular-mean estimator (Eq. 12) makes a window with two distinct lights collapse to one blended point; a natural test is whether replacing it with a mixture of von Mises components on each window histogram recovers the two chromaticities and improves boundary-region accuracy.","The integral-histogram trick is not specific to chroma bins: the same prefix-sum construction could accelerate other local ISP statistics, such as local exposure or local contrast histograms, in the same linear-time regime.","The current window/overlap trade-off means accuracy and boundary sharpness pull in opposite directions; an adaptive window that shrinks near detected illumination edges could combine the low error of large windows with the edge clarity of small ones.","If the CPU timing transfers to mobile DSPs with memory-bandwidth constraints, local AWB could become part of the default viewfinder path; whether it holds under hardware quantization is not tested in the paper."],"forward_implications":["A 64x48 preview image can be white-balanced in 5.8 ms on a CPU, which is inside the 10 ms budget typical of real-time preview pipelines.","With about 0.012M parameters versus 6.4M for AID, IFFCC fits in a footprint where a pixel-level network would not, making local AWB feasible on embedded camera hardware.","Because the features are log-chroma histograms rather than camera-specific learned filters, the same trained model transfers across Canon, Nikon, Sony, and other cameras without per-device fine-tuning, as the Shadow dataset results show.","Spatial smoothing by interpolation and guided filtering gives a continuous illumination map, not per-patch blocks, so video frames avoid abrupt white-balance jumps while light edges remain sharp."],"supporting_citations":[{"why":"Supplies the log-chroma histogram and Fourier-based convolution foundation that IFFCC localizes and accelerates.","marker":"[3]"},{"why":"Provides the integral histogram construction that lets any region's histogram be extracted with a few arithmetic operations.","marker":"[40]"},{"why":"Introduces the log-chromaticity color-localization formulation and convolutional-kernel approach that FFCC and IFFCC build on.","marker":"[2]"},{"why":"Provides the Bivariate von Mises distribution used to read a single mean direction from each circular histogram.","marker":"[38]"},{"why":"Defines the LSMI multi-illuminant benchmark and the pixel-level comparisons used for accuracy evaluation.","marker":"[31]"},{"why":"AID is the pixel-level slot-attention model whose accuracy and parameter count IFFCC is compared against.","marker":"[32]"},{"why":"Guided filtering is the edge-preserving smoother that turns patch estimates into the final illumination map.","marker":"[25]"},{"why":"The Shadow dataset supplies the two-illuminant, multi-camera benchmark used for generalization and ablation experiments.","marker":"[13]"}],"fun_headline_variants":["IFFCC: Multi-illuminant white balance at neural accuracy, 400x leaner","Local AWB: 400x fewer params, 20-100x faster, matching deep nets","Integral histograms + FFT: real-time multi-illuminant white balance","Multi-illuminant AWB in 0.03s CPU: no neural net needed","Speed of FFCC, accuracy of neural nets: IFFCC for multi-illuminant"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The estimate for each window is a single circular mean of that window's filtered histogram, so any window containing two or more clearly distinct lights is represented by one blended point; if that mean falls between the true chromaticities, sharp light boundaries in the final map are inherently blurred.","fun_headline_variants_meta":{"raw":{"variants":["IFFCC: Multi-illuminant white balance at neural accuracy, 400x leaner","Local AWB: 400x fewer params, 20-100x faster, matching deep nets","Integral histograms + FFT: real-time multi-illuminant white balance","Multi-illuminant AWB in 0.03s CPU: no neural net needed","Speed of FFCC, accuracy of neural nets: IFFCC for multi-illuminant"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000321,"raw_usage":{"total_tokens":1826,"prompt_tokens":983,"completion_tokens":843,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":599,"completion_tokens_details":{"reasoning_tokens":722}},"tokens_in":599,"tokens_out":843,"duration_ms":6831,"temperature":1.0,"reasoning_tokens":722,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T10:47:45.800019+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a synthetic two-illuminant image with a straight, sharp boundary and well-separated chromaticities, inspect each window's predicted illuminant against the two known ground-truth lights: if the per-window circular mean (Eq. 12) repeatedly falls between the two true chromaticities rather than near one of them, or if the reconstructed map's angular error on the boundary strip stays well above pixel-level networks, then the local-estimate assumption is the limiting factor.","supporting_citations":[{"cited_title":"Fast fourier color constancy","cited_arxiv_id":null,"evidence_quote":"Supplies the log-chroma histogram and Fourier-based convolution foundation that IFFCC localizes and accelerates."},{"cited_title":"Integral histogram: A fast way to extract histograms in cartesian spaces","cited_arxiv_id":null,"evidence_quote":"Provides the integral histogram construction that lets any region's histogram be extracted with a few arithmetic operations."},{"cited_title":"Convolutional color constancy","cited_arxiv_id":null,"evidence_quote":"Introduces the log-chromaticity color-localization formulation and convolutional-kernel approach that FFCC and IFFCC build on."},{"cited_title":"Statistics of directional data","cited_arxiv_id":null,"evidence_quote":"Provides the Bivariate von Mises distribution used to read a single mean direction from each circular histogram."},{"cited_title":"Large scale multi-illuminant (lsmi) dataset for developing white balance algorithm under mixed illumination","cited_arxiv_id":null,"evidence_quote":"Defines the LSMI multi-illuminant benchmark and the pixel-level comparisons used for accuracy evaluation."},{"cited_title":"Attentive illumination decomposition model for multi-illuminant white balancing","cited_arxiv_id":null,"evidence_quote":"AID is the pixel-level slot-attention model whose accuracy and parameter count IFFCC is compared against."},{"cited_title":"Guided im- age filtering","cited_arxiv_id":null,"evidence_quote":"Guided filtering is the edge-preserving smoother that turns patch estimates into the final illumination map."},{"cited_title":"Shadows & lumination: Two- illuminant multiple cameras color constancy dataset","cited_arxiv_id":null,"evidence_quote":"The Shadow dataset supplies the two-illuminant, multi-camera benchmark used for generalization and ablation experiments."}],"review_version":1}