{"id":"0781b02f-db29-49fc-8ca6-e78e2f87fb69","arxiv_id":"1908.08331","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Adding a Green's function convolution layer that integrates gradient features into spatial features improves saliency detection accuracy and training stability of HED and DSS networks.","lead":"The paper adds a fixed, parameter-free Green's function layer to saliency networks that integrates edge-like features into region-like features. It reports small but consistent F-measure gains on standard benchmarks and improved robustness to noise and low brightness.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The Poisson-decomposition premise is unverified: sideX_3 is never constrained to emit gradient channels, so the GIS gains could come from extra parameters or low-pass filtering, not from Green's-function edge filling.","rationale":"The reader's weakest assumption and my load-bearing concern coincide: Section 2.2.2 admits that nothing forces S and (Gx,Gy) to form a Poisson-consistent decomposition. My attack sharpens why this is load-bearing. Because the GFC is a fixed linear operator that is surjective up to constants on the space of residual maps, the end-to-end loss can be satisfied with arbitrary S and G channels; the GIS output is then effectively a learned fixed nonlinearity rather than a principled edge-filling operation. Under this hypothesis, the consistent F-measure gains in Tables 2-3 could be due to the extra sideX_3 parameters (3 outputs vs 1), the low-pass character of integration, or improved training dynamics, rather than the proposed gradient-domain edge extrapolation. This does not falsify the empirical improvement: the same-seed, same-split comparison is genuine evidence that the architecture change helps HED/DSS, and the repeatability curves and robustness experiments are plausible. It does mean the central mechanistic claim of the paper is not yet established, which is exactly the kind of condition a CONDITIONAL verdict should carry. Since the reader already assigned CONDITIONAL and identified this premise, my read does not move the verdict; it reinforces it. I also note the paper's 'no additional parameters' claim (Sec 5.7) is inaccurate because sideX_3 output channels change from 1 to 3, a small but nonzero parameter increase that should be corrected or clarified in revision.","tokens_in":20730,"tokens_out":9141,"duration_ms":99742,"concrete_test":"On a trained HED-GIS/DSS-GIS, for every image in a held-out validation set and for each side output, extract the three sideX_3 channels S, Gx, Gy before the GIS layer. Apply the same fixed GFC operator of Eq. (9) to (Gx,Gy) to obtain R, then fit the free integration constant and an affine gain (a*R+b) to S. Report the median absolute residual and the R^2 between R_aff and S, compared with a constant baseline. If R_aff explains no more variance in S than a constant map, the learned channels do not satisfy the Poisson relation, and the claimed edge-filling mechanism is not the cause of the observed gains.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.2.2 concedes the key premise: the network is 'never forced, via an intermediate loss, to learn the gradient of the saliency map,' and the gradient channels 'are not necessarily the gradient of the saliency, although they are expected to be.' The training loss only supervises the final GIS output, so the split of sideX_3 into S, Gx, Gy is underdetermined: for any desired output O and any chosen S, the two gradient channels can encode the residual O-S (e.g., choose G = grad(O-S), which the GFC integrates). The claimed mechanism, extrapolating sparse edges into salient regions, is therefore not demonstrated by the F-measure gains. The controlled comparison is also not a pure GIS ablation: sideX_3 is changed from 1 output to 3 outputs, so each 1x1 conv at those layers has three times as many weights (roughly 3,600 extra parameters), contrary to the 'no additional parameters' wording in Sec 5.7. The GFC is also a strong low-pass/integration operator, which could improve robustness or training stability independently of any learned gradient-domain representation. The only evidence for the decomposition is Fig 4, a single illustrative set of feature maps.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a gradient integration and sum (GIS) layer for saliency detection networks. The layer takes three inputs from the final side-output layer of a fully convolutional network: one spatial saliency-like channel and two gradient-domain channels. The two gradient channels are integrated by solving a Poisson equation via a Green's function convolution (GFC) in the Fourier domain, and the integrated result is added to the spatial channel to form the final saliency map. The layer has no learned weights. The authors apply this layer to HED and DSS by changing the final side-layer convolution sideX_3 from one output channel to three output channels. They report improved F-measure and AUC on ECSSD, DUT-OMRON, and PASCAL-S when training on MSRA10K, improved robustness to noise and low brightness, lower sensitivity to initialization and reduced overfitting, and a small computational overhead of about 10 ms. They also report a large improvement when training on DUTS and comparing with AFNet. The central claim is that GFC extrapolates edge features into salient regions inside the network, enabling regional learning without additional parameters.","tokens_in":20969,"tokens_out":3640,"duration_ms":41237,"significance":"If the central claim is correct, the paper makes a useful contribution: introducing a fixed, parameter-free Poisson-integrator layer inside a CNN is novel, and the reported gains on multiple datasets and architectures, along with the robustness to noise and low brightness, could make the GIS layer a simple and broadly applicable plug-in component for saliency models. The controlled comparisons with identical random seeds in Section 5.1 are a genuine strength, as is the authors' candid acknowledgment of failure cases and of the fact that the gradient channels are not forced to be true gradients. However, the paper's mechanistic interpretation, the 'no additional parameters' claim, and the cross-dataset comparison all need substantial additional support before the results can be taken at face value.","major_comments":[{"comment":"The paper's core scientific claim is that the GIS layer improves saliency by extrapolating edge features into salient regions through the Green's function integration. The authors explicitly concede in Sec. 2.2.2 that the network is 'never forced, via an intermediate loss, to learn the gradient of the saliency map' and that the gradient channels 'are not necessarily the gradient of the saliency.' Since the training loss supervises only the final GIS output, the decomposition of sideX_3 into S and (Gx,Gy) is underdetermined: for any target output O and any chosen S, the gradient channels can encode roughly the residual O-S without ever learning a meaningful Poisson relation. The reported F-measure improvements therefore do not by themselves demonstrate the proposed edge-filling mechanism; the same gains could in principle come from the extra capacity in the widened sideX_3 layer or from the strong low-pass/integration behavior of the GFC. The authors should provide direct evidence for the decomposition, for example by measuring the integrability (curl) of the learned gradient channels on held-out data, by ablating the GFC against a generic smoothing/integration operation such as a large Gaussian filter, or by adding an auxiliary loss that enforces Gx and Gy to approximate the gradient of S. Fig. 4, a single illustrative feature-map example, is not sufficient evidence for a load-bearing mechanism.","section":"Sec. 2.2.2 and Eqs. (1)-(9)"},{"comment":"The abstract and Sec. 5.7 state that the GIS layer adds 'no additional parameters' and requires 'very minor architectural changes.' This is inaccurate: replacing the final side-layer output from 1 channel to 3 channels changes the 1x1 convolution at sideX_3 from C_in weights to 3*C_in weights. Based on the channel counts in Table 1, this adds about 3,584 parameters for the six side layers. The number is small relative to VGG-16, but it is not zero, and it changes the capacity of the network at exactly the point where the GIS layer is inserted. This confounds the attribution of the improvement to the GFC operation itself. The authors should either correct the wording, or report the parameter counts for both models and discuss whether the gain can be separated from the widened layer.","section":"Sec. 5.7 and Table 1"},{"comment":"The DUTS comparison is confounded. The authors state that the DSS model 'performs much worse with the DUTS training set' because its hyperparameters were optimized for MSRA10K and 'a new set of parameters need to be chosen,' yet they compare this un-tuned DSS against DSS-GIS using the same MSRA10K hyperparameters. The reported average improvement of 7.4% in F-measure on DUTS may therefore reflect DSS being suboptimally trained rather than a benefit of the GIS layer. To support the claim that GIS improves robustness to training conditions, the authors should tune DSS on the DUTS validation set with a comparable hyperparameter search, or at minimum report DSS results with a few different learning rates and show that DSS-GIS is consistently better across the reasonable hyperparameter range.","section":"Sec. 5.3.2 and Table 5"},{"comment":"All reported numeric results come from a single training run per configuration. The identical-seed design controls for initialization in the paired comparisons, but it does not quantify run-to-run variability, and Fig. 7 shows that DSS training can be quite variable. Several improvements in Tables 2 and 3 are small (0.3-0.6 F-measure), and without multiple seeds or confidence intervals it is not possible to assess whether those gains are statistically meaningful. The authors should report mean and standard deviation over at least three seeds for the main comparisons, or provide a bootstrap analysis on the test set.","section":"Sec. 5.1, Tables 2-7"}],"minor_comments":[{"comment":"The equations and surrounding text contain numerous missing mathematical symbols in the displayed version (e.g., the vector field, the Laplacian, the Green's function), which makes the GFC derivation very hard to follow. Please ensure all glyphs render correctly in the final version.","section":"Sec. 2.1.1, Eqs. (1)-(9)"},{"comment":"The text contains several typos and grammatical errors, including 'lightning changes,' 'it's results,' 'a the brightness is reduced,' and 'the network to operate in the feature domain and in the gradient domain at the same time, thus improving the regional representation via edge filling.' A careful proofread is needed.","section":"Abstract and Sec. 5.4"},{"comment":"The runtime comparison reports 0.08s vs 0.09s and 'about 10ms' additional computation, but it is not stated whether this is on GPU or CPU, the image resolution, or the number of repetitions. Please specify the experimental setup for the timing measurement.","section":"Sec. 5.5"},{"comment":"The paper does not mention whether code or trained models will be released. Given the importance of the controlled comparisons for the central claim, a reproducibility statement about code and data would be valuable.","section":"Sec. 5.7"},{"comment":"The claim that the learned S and (Gx,Gy) channels correspond to saliency-like and gradient-like features is supported by only one illustrative example. Consider providing quantitative evidence across a validation set, for example the average correlation between GFC(Gx,Gy) and S or with the ground-truth saliency.","section":"Fig. 4"}],"recommendation":"major_revision","confidential_remarks":"This is a borderline case. The core idea is interesting and the paired-seed controlled experiments are a positive sign, and I do not think the paper should be rejected outright. However, the mechanical interpretation is not yet supported, the 'no additional parameters' claim is factually incorrect, and the DUTS comparison is confounded. I recommend a major revision that either provides the requested evidence or substantially weakens the mechanistic claims. The authors should also be asked to add multi-seed results so that the small F-measure gains can be assessed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about arXiv:1908.08331. First, the core idea is genuinely new: a fixed Green's function convolution layer (GFC) inserted inside a CNN that integrates gradient-domain features, plus a sum with the spatial features. Second, the paper overstates what the experiments prove. The controlled comparisons are decent, but the main mechanism is unverified and the \"no additional parameters\" claim is simply wrong as written.\n\nWhat's good: The GIS layer is parameter-free and model-agnostic, and the authors show consistent small F-measure gains with identical seeds and the same training procedure for HED and DSS, plus robustness improvements to noise and low brightness. That is worth taking seriously. The GFC math is standard, and the self-citations to prior work on the solver are legitimate.\n\nThe soft spots are real. Section 2.2.2 explicitly says the network is \"never forced\" to learn the gradient of the saliency map, and the Gx/Gy channels \"are not necessarily the gradient of the saliency.\" So the claim that the improvement comes from edge filling is not demonstrated. The gains could come from the extra parameters: sideX_3 goes from 1 to 3 output channels, so each 1x1 conv triples its weights. The paper says \"no additional parameters,\" but that is false. That also means the ablation is not a pure test of the GIS operation. The GFC is also a low-pass integration operator, which could improve robustness independently of any learned gradient decomposition. Only one illustrative feature map is shown (Fig 4), which is not enough evidence. The DUTS comparison is also confounded because the authors keep the MSRA10K hyperparameters for both models, and DSS is known to be sensitive to these; the 7.4% F-measure gain should not be taken at face value. No code or data are released, and final numbers have no error bars.\n\nThese are fixable with a serious revision: release code, add an ablation that matches parameter counts, and include a low-pass control. If the authors can show via auxiliary losses or more extensive visualizations that the network actually learns a Poisson-consistent decomposition, the paper would be much stronger. But as it stands, the central mechanism is a plausible hypothesis, not a demonstrated one.\n\nThis paper deserves a serious referee—it is novel and the experiments are substantial—but the referee will need to push for these corrections. I would bring it to a reading group to discuss the ablation issues, but I would not cite it as solid evidence until the code and ablations are out.","headline":"A genuinely new GIS layer that deserves serious review, but the paper's central mechanism is unverified and the 'no additional parameters' claim is wrong.","tokens_in":21539,"tokens_out":4096,"would_cite":false,"duration_ms":40882,"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":"Adding a fixed, parameter-free edge-filling layer inside a saliency CNN improves F-measure by 1.6% on DUT-OMRON and makes training more repeatable.","keywords":["salient object detection","Green's function convolution","gradient integration and sum","edge extrapolation","Poisson equation","deep convolutional networks","training stability","noise robustness"],"falsifier":"On a trained DSS-GIS, compare the two gradient inputs $\\mathcal{G}_x, \\mathcal{G}_y$ with the numerical gradient of the spatial input $S$ at each side layer, and measure the residual of the Poisson reconstruction $S \\approx \\nabla^{-2}(\\partial_x\\mathcal{G}_x + \\partial_y\\mathcal{G}_y)$. If the residual is large, or if replacing the Green's function convolution with a fixed random convolution of identical cost does not change the test F-measure, the claimed integration mechanism is not what is causing the improvement.","tokens_in":20496,"feed_emoji":"🎯","tokens_out":8072,"duration_ms":78199,"temperature":0.7,"pith_summary":"The paper argues that saliency CNNs suffer because small convolutional kernels can detect edges but cannot turn those edges into filled regions, so it inserts a fixed, parameter-free layer that integrates edge-like gradient features into region-like saliency features. The layer, called gradient integration and sum (GIS), uses a Green's function convolution to solve the implied Poisson equation, effectively giving each pixel information from the whole image at no learned cost. On the HED and DSS architectures, adding GIS at every side output raises the F-measure, most notably an absolute 1.6% gain on DUT-OMRON for DSS-GIS, and makes training faster, more repeatable, and less sensitive to initialization. The same layer improves robustness substantially when test images are degraded by 30% salt-and-pepper noise or an 80% brightness reduction. The improvement comes from a mechanism, gradient integration, rather than from more parameters or a new network family.","feed_headline":"Edge-filling layer lifts saliency accuracy by 1.6 percent","feed_subtitle":"A weight-free Green's function convolution turns detected edges into regions, improving F-measure and resisting noise and low light.","key_machinery":"The GIS layer is built on the Green's function convolution (GFC): the numerical Green's function of the discrete Laplacian, computed once in the Fourier domain as $\\mathcal{V}_{\\mathrm{mono}}^{\\mathcal{F}} = \\mathcal{F}(\\delta_p) / \\mathcal{F}(\\nabla_p^2)$, is applied as a fixed convolution to integrate any gradient-like vector field. The layer takes three inputs per side output, $S$ in the spatial domain and $\\mathcal{G}_x, \\mathcal{G}_y$ in the gradient domain, computes the Laplacian $\\Delta = \\partial_x \\mathcal{G}_x + \\partial_y \\mathcal{G}_y$, solves the Poisson equation by convolution with the Green's function, and sums the integrated field with $S$. No weights are learned; the network learns only which three channels to emit. The mechanism's work is to give the network a whole-image receptive field for free, so thin edge features from high-resolution side layers can be extrapolated into smooth filled regions rather than requiring large learned kernels or downscaling.","core_discovery":"The central claim is that a saliency CNN's side outputs can be split into three channels—one spatial saliency-like channel and two gradient-domain channels—and that integrating the gradient channels with the numerical Green's function of the discrete Laplacian, then adding the result to the spatial channel, produces a better saliency map than the network's original output. The authors call this the GIS layer, and they add it without weights, without an intermediate loss, and with only about 10 ms of extra computation. In their experiments, DSS-GIS outperforms the paper's DSS implementation on all three test sets, with the largest gains on DUT-OMRON, where the F-measure rises by 1.6%, and HED-GIS lifts HED from a poor saliency model to near the level of saliency-focused networks. They also report that the layer reduces overfitting, narrows the spread of validation curves across random initializations, and yields larger margins under noise, with a 5.2% F-measure gain on DUT-OMRON under 30% salt-and-pepper noise and 2.8% under an 80% brightness reduction.","pith_inferences":["The mechanism claim implies a testable signature: if the learned gradient channels truly approximate the gradient of the spatial saliency channel, the Poisson residual in a trained DSS-GIS should be small. The paper does not report such a residual check.","Since the GFC gives an unlimited receptive field with zero parameters, placing GIS deeper inside the backbone, before later convolutions collapse resolution, could yield larger gains than the final-layer placement tested here; the paper lists this as future work.","The robustness results suggest the gains may be larger precisely where texture and contrast cues are unreliable, so standard clean-image benchmarks may understate the layer's value compared with real low-light or noisy capture.","The comparison with a newer saliency architecture suggests a fixed geometric prior can rival architectural advances on standard benchmarks; whether that holds on very recent networks is an open question the paper does not test."],"forward_implications":["Any fully convolutional saliency network can receive the GIS layer at its side outputs with no architectural redesign and no increase in parameter count, so the reported gains are not tied to HED or DSS specifically.","Training becomes more repeatable and less sensitive to initialization and hyperparameters, which should make saliency models easier to deploy in new training regimes.","GIS can replace denseCRF post-processing with a 40 times faster and usually more accurate in-network alternative, and the paper reports that combining the two yields the best results.","Degraded-input robustness suggests saliency systems using GIS will hold up better in low-light or noisy imagery without retraining.","HED-GIS turns a network that originally performed poorly at saliency into a competitive saliency model, showing that edge extrapolation can substitute for an architecture designed for regions."],"supporting_citations":[{"why":"Supplies the Green's function convolution solver that the GIS layer uses as its integration step, including padding and Fourier-domain kernel construction.","marker":"[20]"},{"why":"Defines the HED side-output architecture that the paper modifies by adding a GIS layer at each side output.","marker":"[2]"},{"why":"Defines the DSS architecture with short connections that the paper extends into DSS-GIS and reimplements as the baseline.","marker":"[4]"},{"why":"Provides denseCRF, the post-processing method used as the main comparison for the GIS layer's speed and accuracy.","marker":"[15]"},{"why":"Provides the hardest of the three test sets, where the largest F-measure gain of 1.6% is reported.","marker":"[12]"},{"why":"Supplies the training set for the main experiments, split into training, validation, and test partitions.","marker":"[34]"},{"why":"Supplies the recent baseline trained on the newer training set that DSS-GIS is compared against.","marker":"[40]"},{"why":"Supplies the Adam optimizer used in the training procedure that the paper changes from the original DSS setup.","marker":"[33]"},{"why":"Supplies the newer training and test sets used to show the model adapts without changing hyperparameters.","marker":"[38]"}],"fun_headline_variants":["Green's function layer fills edges to boost saliency by 1.6%","Weight-free edge-filling layer improves saliency and noise robustness","GIS layer: edge integration lifts saliency F-measure by 1.6%","Saliency CNN gains from Green's function convolution","Edge-filling convolution sharpens saliency, resists noise and dim light"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the three output channels of each side layer will spontaneously organize, under only the final saliency loss, into one region-like channel and two channels that behave like the gradient of that region; no intermediate loss or constraint enforces this split.","fun_headline_variants_meta":{"raw":{"variants":["Green's function layer fills edges to boost saliency by 1.6%","Weight-free edge-filling layer improves saliency and noise robustness","GIS layer: edge integration lifts saliency F-measure by 1.6%","Saliency CNN gains from Green's function convolution","Edge-filling convolution sharpens saliency, resists noise and dim light"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000527,"raw_usage":{"total_tokens":2635,"prompt_tokens":1128,"completion_tokens":1507,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":744,"completion_tokens_details":{"reasoning_tokens":1412}},"tokens_in":744,"tokens_out":1507,"duration_ms":11733,"temperature":1.0,"reasoning_tokens":1412,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:41:24.670588+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a trained DSS-GIS, compare the two gradient inputs $\\mathcal{G}_x, \\mathcal{G}_y$ with the numerical gradient of the spatial input $S$ at each side layer, and measure the residual of the Poisson reconstruction $S \\approx \\nabla^{-2}(\\partial_x\\mathcal{G}_x + \\partial_y\\mathcal{G}_y)$. If the residual is large, or if replacing the Green's function convolution with a fixed random convolution of identical cost does not change the test F-measure, the claimed integration mechanism is not what is causing the improvement.","supporting_citations":[],"review_version":1}