Pith. sign in

REVIEW 4 major objections 5 minor 3 references

Deep Green Function Convolution for Improving Saliency in Convolutional Neural Networks

T0 review · 4 major / 5 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read 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.

desk verdict 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. read the letter →

arxiv 1908.08331 v2 pith:SGLDSRFT submitted 2019-08-22 cs.CV cs.LG

Dominique Beaini , Sofiane Achiche , Alexandre Duperré , Maxime Raison This is my paper · ORCID
classification cs.CVcs.LG
keywords salientobjectdetectionGreen'sfunctionconvolutiongradientintegrationandsumedgeextrapolationPoissonequationdeepconvolutionalnetworkstrainingstabilitynoiserobustness
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

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

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

  • 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.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

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.

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 (4)
  1. [Sec. 2.2.2 and Eqs. (1)-(9)] 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.
  2. [Sec. 5.7 and Table 1] 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.
  3. [Sec. 5.3.2 and Table 5] 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.
  4. [Sec. 5.1, Tables 2-7] 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.
minor comments (5)
  1. [Sec. 2.1.1, Eqs. (1)-(9)] 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.
  2. [Abstract and Sec. 5.4] 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.
  3. [Sec. 5.5] 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.
  4. [Sec. 5.7] 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.
  5. [Fig. 4] 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.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the GIS layer is a fixed, parameter-free module evaluated on external benchmarks with controlled training.

full rationale

The central claim is that adding a fixed GIS layer to HED/DSS improves saliency F-measure. That claim is an empirical result: the authors train HED/DSS with and without the GIS layer using the same code, random seed, data split, and training procedure, and then evaluate on external benchmarks (ECSSD, DUT-OMRON, PASCAL-S, DUTS). The GIS layer itself contains no learned parameters; its Green's function convolution kernel is derived in Eqs. (1)-(9) from the standard numerical Laplacian and the convolution theorem, so no quantity is fitted to the reported test metrics. The only fitted objects are the ordinary network weights, trained end-to-end with the same cross-entropy loss whether or not the GIS layer is present. The self-citations to Beaini et al. [19,20] provide the GFC solver and the edge-filling property, but the paper restates the solver equations and the benchmark gains stand independently of those citations. The admission in Sec. 2.2.2 that the network is never forced to learn a true gradient decomposition is a limitation of the explanatory mechanism, not a circularity: the improvement claim does not reduce by construction to that assumption. Any inaccuracy in the 'no additional parameters' wording (sideX_3 changes from 1 to 3 output channels) is a correctness concern, not circularity.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The method relies on a fixed Poisson-solving layer with no learned parameters. The free parameters listed are training and numerical choices that affect results but are not fitted to the target claim. The axioms capture the mathematical foundation from prior work and the empirical assumption that the 3-channel decomposition emerges naturally during training.

free parameters (4)
  • Learning rate = 4e-5 (Adam)
    Chosen by cross-validation in Sec 4.2; identical across compared models.
  • Mini-batch size = 8
    Changed from DSS original 10; chosen by hand in Sec 2.2.3.
  • GFC padding = 4 pixels
    Padding value 0 added around the Laplacian to avoid discontinuities; from Sec 2.1.1 and [20].
  • Training iterations = 30k plus 2k fine-tune
    Set in Sec 2.2.3 with early stopping on validation.
assumptions (3)
  • standard math The Green's function convolution (GFC) solves the discrete Poisson equation with minimal error when applied in the Fourier domain.
    Taken from prior work [20]; used in Eqs. (3)-(9) to justify the integration step.
  • domain assumption The sideX_3 layers can learn the S / gx / gy decomposition without an intermediate loss.
    Sec 2.2.2 states the network is never forced to learn the gradient of the saliency; the improvement relies on this emergent decomposition.
  • ad hoc to paper The fused output S + GFC(gx, gy) is a meaningful saliency enhancement rather than an arbitrary nonlinearity.
    This is the paper's core hypothesis; empirical validation in Sec 4-5, but no theoretical justification.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Deep Green Function Convolution for Improving Saliency in Convolutional Neural Networks." pith.science (2026). https://pith.science/paper/SGLDSRFT

@misc{pith2026190808331,
  author       = {Pith},
  title        = {Pith review of: Deep Green Function Convolution for Improving Saliency in Convolutional Neural Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SGLDSRFT}},
  note         = {Machine review of arXiv:1908.08331}
}
read the original abstract

Current saliency methods require to learn large scale regional features using small convolutional kernels, which is not possible with a simple feed-forward network. Some methods solve this problem by using segmentation into superpixels while others downscale the image through the network and rescale it back to its original size. The objective of this paper is to show that saliency convolutional neural networks (CNN) can be improved by using a Green's function convolution (GFC) to extrapolate edges features into salient regions. The GFC acts as a gradient integrator, allowing to produce saliency features by filling thin edges directly inside the CNN. Hence, we propose the gradient integration and sum (GIS) layer that combines the edges features with the saliency features. Using the HED and DSS architecture, we demonstrated that adding a GIS layer near the network's output allows to reduce the sensitivity to the parameter initialization, to reduce the overfitting and to improve the repeatability of the training. By simply adding a GIS layer to the state-of-the-art DSS model, there is an absolute increase of 1.6% for the F-measure on the DUT-OMRON dataset, with only 10ms of additional computation time. The GIS layer further allows the network to perform significantly better in the case of highly noisy images or low-brightness images. In fact, we observed an F-measure improvement of 5.2% when noise was added to the dataset and 2.8% when the brightness was reduced. Since the GIS layer is model agnostic, it can be implemented into different fully convolutional networks. A major contribution of the current work is the first implementation of Green's function convolution inside a neural network, which allows 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.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

3 extracted references · 2 canonical work pages

  1. [549]

    https://doi.org/10.1109/CVPR.2017.65

    IEEE, Honolulu, HI (2017). https://doi.org/10.1109/CVPR.2017.65. 18. Kuen, J., Wang, Z., Wang, G.: Recurrent Attentional Networks for Saliency Detection. In: 2016 IEEE C onference on Computer Vision and Pattern Recognition (CVPR). pp. 3668 –

    work page doi:10.1109/cvpr.2017.65 2017

  2. [3677]

    Computing the Spatial Probability of Inclusion inside Partial Contours for Computer Vision Applications

    IEEE, Las Vegas, NV, USA (2016). https://doi.org/10.1109/CVPR.2016.399. 19. Beaini, D., Achiche, S., Nonez, F., Raison, M.: Computing the Spatial Probability of Inclusion inside Pa rtial Contours for Computer Vision Applications. arXiv:1806.01339 [cs, math]. (2018). 20. Beaini, D., Achiche, S., Nonez, F., Brochu Dufour, O., Leblond - Ménard, C., Asaadi, M...

    work page Pith review arXiv 2016

  3. [3805]

    https://doi.org/10.1109/CVPR.2017.404

    IEEE, Honolulu, HI (2017). https://doi.org/10.1109/CVPR.2017.404. 39. The DUTS Image Dataset, http://saliencydetection.net/duts/, last accessed 2019 /10/31. 40. Feng, M., Lu, H., Ding, E.: Attentive Feedback Network for Boundary - Aware Salient Object Detection. 10. 41. Bylinskii, Z., Judd, T., Oliva, A., Torralba, A., Durand, F.: What do different evalua...

    arXiv 2017

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.