Pith. sign in

REVIEW 4 major objections 5 minor 32 references

See More Than Once -- Kernel-Sharing Atrous Convolution for Semantic Segmentation

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

Pith's one-line read The paper claims that sharing one 3x3 kernel across parallel atrous branches at different rates lets a segmentation network see each feature map multiple times, improving accuracy while reducing parameters, and that this shared-kernel…

desk verdict Novel kernel-sharing atrous module with a clear efficiency gain, but the reported accuracy improvement is confounded with parameter reduction and the stated training-sample mechanism does not hold up. read the letter →

arxiv 1908.09443 v4 pith:MFQULLFE submitted 2019-08-26 cs.CV

classification cs.CV
keywords semanticsegmentationatrousconvolutiondilatedkernelsharingmulti-scalecontextASPPPASCALVOC2012ADE20K
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 tries to show that a single 3x3 kernel, shared across parallel atrous-convolution branches with different dilation rates, can outperform the standard ASPP module in semantic segmentation while using far fewer parameters. The idea is that the same kernel sees each feature map multiple times at different receptive fields, so the network gains multi-scale context and every object contributes to training one kernel rather than one branch-specific kernel. On PASCAL VOC 2012 with Xception, KSAC raises mIOU from 83.34% to 85.96% and saves about 10M parameters; with rates extended to (1,6,12,18,24), mIOU reaches 87.01%. If correct, KSAC offers a simple drop-in replacement for ASPP that is smaller, faster, and better able to exploit wider context.

What carries the argument

The central object is the kernel-sharing atrous convolution module: one 3x3 kernel reused by parallel atrous branches at different rates (typically 6, 12, 18, and optionally 1 and 24), with batch normalization and ReLU applied to each branch output before concatenation together with image-level features. This carries the argument because it makes the module's parameter count constant in the number of branches rather than linear, and because every object in the training images contributes to training the single shared kernel at every receptive field.

What would settle it

Train a comparison model in which each atrous branch keeps its own 3x3 kernel but with output channel count reduced so total parameters match KSAC; if this model matches or exceeds KSAC's mIOU, the gain is not caused by sharing the kernel.

Watch

Extended reading notes

Core claim

The paper's central claim is that replacing the parallel atrous branches of ASPP, each with its own 3x3 kernel, with a single 3x3 kernel shared across the same branches improves semantic segmentation accuracy and reduces model size. The shared kernel sees the same feature map at multiple atrous rates, which the paper argues increases each kernel's effective training samples and lets it learn both local detail and global context. On PASCAL VOC 2012 validation with an Xception backbone, KSAC raises mIOU from 83.34% to 85.96% while cutting parameters from 54.3M to 44.8M; extending the rates to (1,6,12,18,24) raises mIOU further to 87.01% without adding parameters. On the VOC test set KSAC reaches 88.1% mIOU, and on ADE20K validation it reaches 45.47% mIOU under multi-scale evaluation.

Load-bearing premise

The load-bearing premise is that the accuracy gains come from the sharing mechanism itself rather than from the accompanying reduction in parameters or other differences between the compared setups.

Editorial extensions

If this is right

  • With the Xception backbone on PASCAL VOC 2012 validation, replacing ASPP with KSAC raises mIOU from 83.34% to 85.96% and cuts model size from 54.3M to 44.8M parameters.
  • With MobileNetV2, KSAC reduces parameters by about one third (4.5M to 3.0M) while improving mIOU from 75.70% to 76.30%.
  • Adding rates 1 and 24 to KSAC improves mIOU from 85.96% to 87.01% with no additional parameters, whereas ASPP degrades when rate 24 is added.
  • KSAC reports 88.1% mIOU on the PASCAL VOC 2012 test set and 45.47% mIOU on ADE20K validation with multi-scale evaluation, under output stride 16.
  • Because the shared kernel is reused by every branch, adding more atrous branches does not increase the module's parameter count, so wider context can be explored at no memory cost in the convolution weights.

Reading between the lines

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

  • A direct test of the paper's 'increased effective training samples' explanation would be to compare per-class accuracy between KSAC and ASPP: the sharing hypothesis predicts larger gains for classes with extreme object scales, not uniform gains.
  • The sharing mechanism is not specific to segmentation; a similar single-kernel multi-rate design could replace dilated parallel branches in object detection or instance segmentation heads, where multi-scale context also matters.
  • A capacity-matched ASPP variant with reduced per-branch channels would separate the contribution of sharing from the contribution of having fewer parameters, and would clarify whether the mechanism itself or the regularization effect of fewer parameters drives the reported gains.
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 Kernel-Sharing Atrous Convolution (KSAC), a modification of Atrous Spatial Pyramid Pooling (ASPP) in which a single 3x3 kernel is shared across parallel atrous branches with different rates, so that the same kernel processes the input feature maps multiple times with different receptive fields. The authors argue that this sharing improves the generalization and representation ability of the kernels, increases the effective number of training samples, reduces model parameters, and allows the network to benefit from a wider range of atrous rates. Experiments on PASCAL VOC 2012 and ADE20K report consistent mIOU improvements over the ASPP baseline with reduced parameter counts: for example, with an Xception backbone on VOC 2012 val, mIOU rises from 83.34% (DeepLabV3+/ASPP) to 85.96% (KSAC) while parameters drop from 54.3M to 44.8M, and extending the rate set to (1,6,12,18,24) further improves mIOU to 87.01%. The paper also reports 88.1% on the VOC 2012 test set and 45.47% on ADE20K validation.

Significance. If the reported gains are attributable to the kernel-sharing mechanism itself, KSAC is a simple, parameter-efficient drop-in replacement for ASPP that could be broadly useful in semantic segmentation and other dense prediction tasks. The empirical comparisons against public benchmarks are appropriate for the claim, and the reported improvements are consistent across two backbones and a second dataset, which strengthens the plausibility. However, the central empirical claim is not yet isolated from confounding factors: the main comparison changes both the sharing constraint and the total parameter count simultaneously, and the paper provides no ablation that matches capacity while keeping kernels separate, no error bars or multiple-seed results, and no released code. The conceptual explanation in Section 3.2 is asserted rather than measured, and the speed/memory comparison with prior work is not controlled for output stride. Thus the result is promising but the causal attribution to sharing is not yet securely established.

major comments (4)
  1. [Table 1, Section 5.1] The primary comparison between KSAC and ASPP changes at least two variables at once: the weight-sharing constraint and the total number of parameters (54.3M vs 44.8M for Xception). Consequently, the reported mIOU gain could be caused by reduced capacity acting as regularization, by different optimization dynamics, or by implementation differences rather than by the sharing mechanism itself. Please add a capacity-matched control, such as an ASPP variant with separate per-branch 3x3 kernels but output channels per branch reduced so that the total 3x3 parameter budget equals the single shared kernel, and report mIOU and parameter counts for both variants.
  2. [Section 3.2] The claim that sharing increases the number of effective training samples is not supported by the paper's own setup. In ASPP, every training image is already passed through every parallel branch, so every image contributes gradients to each branch kernel; the assertion that small or large objects are only effective for training the branch with a matching atrous rate is not substantiated. Please either remove this explanation or support it with direct evidence, for example gradient statistics per branch, or an ablation that isolates the sharing effect.
  3. [Sections 5.3 and 5.4, Tables 2 and 3] The comparison with prior work is not controlled for output stride: the KSAC results in Table 3 are obtained with OS=16, while the competing results from DeepLab V3, DeepLab V3+, and EMA are obtained with OS=8. The paper also claims in Section 5.4 that KSAC achieves similar segmentation results under OS=8 and OS=16, but no OS=8 experiment is reported anywhere. This weakens both the accuracy comparison and the speed/memory claims, and the relevant experiment or a clear caveat should be added.
  4. [Section 5.1 and Table 1] No error bars, multiple-seed results, or statistical significance tests are reported for any of the mIOU differences. Given that the MobileNetV2 improvement is only 0.6 percentage points (75.70% vs 76.30%), run-to-run and implementation variation could plausibly account for the gap. Please report variance across at least three runs for the key comparisons, or otherwise justify that the differences are stable.
minor comments (5)
  1. [Abstract and Introduction] There are typos and formatting issues, such as 'whe compared' in the abstract and 'Tensorflow2.0' in Section 4.2; these should be corrected.
  2. [Algorithm 1] The notation KERNEL(shape) is not defined; please specify how the shared kernel is initialized and whether it is updated by gradients from all branches jointly.
  3. [Section 5.2] The sentence 'when the Xception decoder is used, about ten times of parameters have been reduced compared with MobileNetV2' is unclear; it likely refers to the Xception encoder, and the magnitude should be stated more precisely.
  4. [Figure 5] The feature-map visualization is descriptive but not quantitative; consider adding a numerical measure, such as gradient statistics or feature discriminability, to support the claim of clearer edges and contours.
  5. [References] Several references are incomplete or inconsistently formatted, for example [14] lists 'International Conference on Computer Vision and Pattern Recognition' where the venue should be ICCV, and some arXiv identifiers appear without full citation details.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's claims are empirical architecture comparisons against public benchmarks, with no quantity fitted to the target metric and no load-bearing self-citation chain.

full rationale

The central claim is that the proposed KSAC module improves semantic segmentation accuracy and reduces parameters relative to ASPP. This is supported by benchmark experiments on PASCAL VOC 2012 and ADE20K using external labels and published baselines. No parameter is fitted to the reported mIOU; the atrous rates and kernel-sharing design are fixed beforehand, and the measured improvements are not derived from the paper's own assumptions. The parameter-count formula in Section 3.2 is an arithmetic consequence of sharing one kernel, not a prediction of mIOU, so it is not circular. The paper's explanation that sharing increases effective training samples and performs feature augmentation is an interpretive claim rather than a definitional one; even if unvalidated, it does not make the headline result circular. References to DeepLab and other prior works are standard baseline citations and are not authored by the present paper's authors, so there is no load-bearing self-citation or imported uniqueness theorem. The absence of a capacity-matched ablation is a substantive validity concern about whether the gain is caused by sharing itself or by parameter reduction, but that is a confounding-variable issue, not a circularity issue: the prediction is not defined in terms of the fitted input. Accordingly, per the rubric, the honest finding is no significant circularity (score 0).

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

This is an empirical architecture paper. No theoretical free constants or invented entities are introduced; the listed items are hand-chosen design choices and background assumptions on which the benchmark comparison depends.

free parameters (2)
  • Atrous rate sets = (6, 12, 18) and (1, 6, 12, 18, 24)
    Hand-chosen from DeepLab's standard setting; no search or sensitivity analysis. The wider-context claim depends on the choice (1,6,12,18,24).
  • Output stride = 16
    All reported models use OS=16, while some compared published methods use OS=8; the efficiency and memory claims rest on this design choice.
assumptions (3)
  • domain assumption PASCAL VOC 2012 and ADE20K mIOU benchmarks are valid, comparable measures of segmentation quality.
    Every conclusion about improvement rests on these public benchmarks; scores are taken as faithful, external evidence.
  • domain assumption The ASPP baseline in Table 1 is a faithful reproduction of DeepLab V3+ under identical code, pretraining, and hyperparameters.
    The central KSAC-versus-ASPP comparison assumes the only difference is the module; no code, training logs, or seeds are supplied to confirm identical conditions.
  • ad hoc to paper Weight tying across atrous branches preserves enough capacity that observed gains can be attributed to the sharing mechanism.
    This is the paper's core hypothesis in Section 3.2; it is asserted rather than proven and is not isolated by an ablation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of See More Than Once -- Kernel-Sharing Atrous Convolution for Semantic Segmentation." pith.science (2026). https://pith.science/paper/MFQULLFE

@misc{pith2026190809443,
  author       = {Pith},
  title        = {Pith review of: See More Than Once -- Kernel-Sharing Atrous Convolution for Semantic Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MFQULLFE}},
  note         = {Machine review of arXiv:1908.09443}
}
read the original abstract

The state-of-the-art semantic segmentation solutions usually leverage different receptive fields via multiple parallel branches to handle objects with different sizes. However, employing separate kernels for individual branches degrades the generalization and representation abilities of the network, and the number of parameters increases linearly in the number of branches. To tackle this problem, we propose a novel network structure namely Kernel-Sharing Atrous Convolution (KSAC), where branches of different receptive fields share the same kernel, i.e., let a single kernel see the input feature maps more than once with different receptive fields, to facilitate communication among branches and perform feature augmentation inside the network. Experiments conducted on the benchmark PASCAL VOC 2012 dataset show that the proposed sharing strategy can not only boost a network s generalization and representation abilities but also reduce the model complexity significantly. Specifically, on the validation set, whe compared with DeepLabV3+ equipped with MobileNetv2 backbone, 33% of parameters are reduced together with an mIOU improvement of 0.6%. When Xception is used as the backbone, the mIOU is elevated from 83.34% to 85.96% with about 10M parameters saved. In addition, different from the widely used ASPP structure, our proposed KSAC is able to further improve the mIOU by taking benefit of wider context with larger atrous rates. Finally, our KSAC achieves mIOUs of 88.1% and 45.47% on the PASCAL VOC 2012 test set and ADE20K dataset, respectively. Our full code will be released on the Github.

Figures

Figures reproduced from arXiv: 1908.09443 by the authors.

Figure 1
Figure 1. The multi-branch-like solutions used in PSPNet [30] and [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Illustration of our proposed Kernel-Sharing Atrous Con [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 4
Figure 4. The detailed architecture of our proposed Kernel-Sharing [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: Visualization of the feature maps extracted by kernels of KSAC and ASPP. Here, [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: Comparison of the segmentation results obtained by [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

32 extracted references · 27 canonical work pages

  1. [1]

    Collins, Yukun Zhu, George Papandreou, et al

    Liang-Chieh Chen, Maxwell D. Collins, Yukun Zhu, George Papandreou, et al. Searching for efficient multi-scale archi- tectures for dense image prediction. NIPS, 2018. 8

  2. [2]

    Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Mulphy, and Alan L. Yuille. Deeplab: semantic im- age segmentation with deep convolutional nets, atrous con- volution and fully connected crfs. IEEE Transaction on Pat- tern Recognition and Machine Intelligence , 40(4):834–848, 2018

  3. [3]

    Semantic image seg- mentation with deep convolutional nets and fully connected crfs

    Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Semantic image seg- mentation with deep convolutional nets and fully connected crfs. International Conference on Learning Representations, 2017

  4. [4]

    Rethinking atrous convolution for semantic image segmentation

    Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv:1706.05587v3, 2017

  5. [5]

    Encoder-decoder with atrous separable convolution for semantic image segmentation

    Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. Eu- ropean Conference on Computer Vision, 2018

  6. [6]

    Stacked deconvolutional network for se- mantic segmentation

    Jun Fu, Jing Liu, Yuhang Wang, Jin Zhou, Changyong Wang, and Hanqing Lu. Stacked deconvolutional network for se- mantic segmentation. IEEE Transaction on Image Process- ing, 2019

  7. [7]

    Ce-net: Context encoder network for 2d medical im- age segmentation

    Zaiwang Gu, Jun Cheng, Huazhu Fu, Kang Zhou, and the others. Ce-net: Context encoder network for 2d medical im- age segmentation. IEEE Transaction on Medical Imaging , 2019

  8. [8]

    Semantic contours from inverse detectors

    Bharath Hariharan, Pablo Arbelaez, Lubomir Bourdev, Subhransu Maji, and Jitendra Malik. Semantic contours from inverse detectors. International Conference on Computer Vi- sion, 2011

Show all 32 references
  1. [9]

    Tao Hu, Pengwan Yang, Chiliang Zhang, Gang Yu, Yadong Mu, and Cees G. M. Snoek. Attention-based multi-context guiding for few-shot semantic segmentation. AAAI Confer- ence on Artificial Intelligence, 2019

  2. [10]

    Expectation-maximization attention net- works for semantic segmentation

    Xia Li, Zhisheng Zhong, Jianlong Wu, Yibo Yang, Zhouchen Lin, and Hong Liu. Expectation-maximization attention net- works for semantic segmentation. International Conference on Computer Vision, 2019

  3. [11]

    Pixel-anchor: a fast oriented scene text detector with combined networks

    Yuan Li, Yuanjie Yu, Zefeng Li, Yangkun Lin, Meifang Xu, Jiwei Li, and Xi Zhou. Pixel-anchor: a fast oriented scene text detector with combined networks. arXiv:1811.07432, 2018

  4. [12]

    Liang, H

    X. Liang, H. Zhou, and E. Xing. Dynamic-structured seman-tic propagation network. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages752761, 2018

  5. [13]

    G. Lin, A. Milan, C. Shen, and I. Reid. Refinenet: Multi-path refinement networks for high-resolution semantic segmenta- tion. Proceedings of the IEEE Conference on Computer Vi- sion and Pattern Recognition, 2017

  6. [14]

    Auto-deeplab: hierarchical neural architecture search for semantic image segmentation

    Chenxi Liu, Liang-Chieh Chen, Florian Schroff, Hartwig Adam, Wei Hua, Alan Yuille, and Li Fei-Fei. Auto-deeplab: hierarchical neural architecture search for semantic image segmentation. International Conference on Computer Vision and Pattern Recognition, 2019

  7. [15]

    Fully convolutional networks for semantic segmentation

    Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. Interna- tional Conference on Computer Vision and Pattern Recogni- tion, pages 3431–3440, 2015

  8. [16]

    U-net: convolutional networks for biomedical image segmentation

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: convolutional networks for biomedical image segmentation. International Conference on Medical Image Computing and Computer-Assisted Intervention, pages 234–241, 2015

  9. [17]

    Mobilenetv2: Inverted residuals and linear bottlenecks

    Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh- moginov, and Liang cheich Chen. Mobilenetv2: Inverted residuals and linear bottlenecks. International Conference on Computer Vision and Pattern Recognition, 2018

  10. [18]

    High-resolution representations for labeling pixels and regions

    Ke Sun, Yang Zhao, Borui Jiang, Tianheng Cheng, Bin Xiao, et al. High-resolution representations for labeling pixels and regions. abs/1904.04514, 2019

  11. [19]

    Self- supervised model adaptation for multimodal semantic seg- mentation

    Abhinav Valada, Rohit Mohan, and Wolfram Burgard. Self- supervised model adaptation for multimodal semantic seg- mentation. International Journal of Computer Vision, 2019

  12. [20]

    Fastfcn: rethinking dilated convolution in the backbone for semantic segmentation

    Huikai Wu, Junge Zhang, Kaiqi Huang, Kongming Liang, and Yizhou Yu. Fastfcn: rethinking dilated convolution in the backbone for semantic segmentation. arxiv:1903.11816, 2019

  13. [21]

    Group normalization

    Yuxin Wu and Kaiming He. Group normalization. European Conference on Computer Vision, 2018

  14. [22]

    T. Xiao, Y . Liu, B. Zhou, Y . Jiang, and J. Sun. Unified per- ceptual parsing for scene understanding. arXiv:1807.10221, 2018

  15. [23]

    Self-ensembling attention attention net- works: addressing domain shift for semantic segmentation

    Yonghao Xu, Bo Du, Lefei Zhang, Qian Zhang, Guoli Wang, and Liangpei Zhang. Self-ensembling attention attention net- works: addressing domain shift for semantic segmentation. AAAI Conference on Artificial Intelligence, 2019

  16. [24]

    Denseaspp for semantic segmentation in street scenes

    Maoke Yang, Kun Yu, Chi Zhang, Zhiwei Li, and Kuiyuan Yang. Denseaspp for semantic segmentation in street scenes. International Conference on Computer Vision and Pattern Recognition, pages 3684–3692, 2018

  17. [25]

    Con- text encoding for semantic segmentation

    Hang Zhang, Kristin Dana, Jianping Shi, Zhongyue Zhang, Xiaogang Wang, Ambrish Tyagi, and Amit Agrawal. Con- text encoding for semantic segmentation. International Con- ference on Computer Vision and Pattern Recognition, 2018

  18. [26]

    Zhang, K

    H. Zhang, K. Dana, J. Shi, Z. Zhang, X. Wang, A. Tyagi, and A. Agrawal. Context encoding for semantic segmenta- tion. The IEEE Conference on Computer Vision and Pattern Recognition, 2018

  19. [27]

    Co-occurrent features in semantic segmentation

    Hang Zhang, Han Zhang, Chenguang Wang, and Junyuan Xie. Co-occurrent features in semantic segmentation. Inter- national Conference on Computer Vision and Pattern Recog- nition, 2019

  20. [28]

    Zhang, S

    R. Zhang, S. Tang, Y . Zhang, J. Li, and S. Yan. Scale- adaptive convolutions for scene parsing.Proc. 26th Int.Conf. Comput. Vis., pages 20312039, 2017, 2017

  21. [29]

    Exfuse: enhancing feature fusion for semantic segmentation

    Zhenli Zhang, Xiangyu Zhang, Chao Peng, Xiangyang Xue, and Jian Sun. Exfuse: enhancing feature fusion for semantic segmentation. European Conference on Computer Vision , 2018

  22. [30]

    Pyramid scene parsing network

    Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. Inter- national Conference on Computer Vision and Pattern Recog- nition, pages 2881–2890, 2017. 9

  23. [31]

    Learning fully dense neural networks for image se- mantic segmentation

    Mingmin Zhen, Jinlu Wang, Lei Zhou, Tian Fang, and Long Quan. Learning fully dense neural networks for image se- mantic segmentation. AAAI Conference on Artificial Intelli- gence, 2019

  24. [32]

    Scene parsing through ade20k dataset

    Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. International Conference on Computer Vi- sion and Pattern Recognition, 2017. 10

Pith tools

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