Pith. sign in

REVIEW 4 major objections 5 minor 15 references

Don't ignore Dropout in Fully Convolutional Networks

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

Pith's one-line read Carefully placed dropout lifts a low-data segmentation model's mIoU from 0.49 to 0.59.

desk verdict A useful but statistically under-powered ablation of dropout variants for low-data segmentation; the headline gain is within sampling noise as reported. read the letter →

arxiv 1908.09162 v1 pith:XTVBBQRM submitted 2019-08-24 cs.CV

classification cs.CV
keywords dropoutspatialsemanticsegmentationlow-datatrainingregularizationDeepLabv3+ScheduledDropPathBatchNormalization
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 dropout is not obsolete in fully convolutional networks, provided it is applied in the right form and the right layer. On a deliberately small training set (10% of PASCAL VOC 2012's segmentation classes), it reimplements DeepLabv3+ and tests dropout variants: channel-wise SpatialDropout, DropBlock, and UOut, with and without a linear dropout schedule. Its headline result is that applying SpatialDropout across the backbone, pyramid pooling, and decoder, with ScheduledDropPath ramping in over 30 epochs, raises mean mIoU from 0.49 to 0.59 relative to the no-dropout baseline. The practical significance is that segmentation labels are expensive, so a training-time-only change that buys a 0.10 accuracy gain under low data would be worth adopting even in architectures that use BatchNorm.

What carries the argument

The load-bearing mechanism is channel-wise dropout, called SpatialDropout: instead of zeroing individual pixels, it zeroes entire feature maps along the channel dimension at training time. Because convolutional activations are spatially correlated, pixel dropout does not remove a feature, while channel dropout removes the whole feature from the image and forces later layers to work from the remaining channels. The second mechanism is ScheduledDropPath, a linear ramp that increases dropout probability from zero to its full value over the first 30 epochs, letting the model fit the data before regularization kicks in. A third piece is the variance-shift argument for why dropout before BatchNorm is harmful, which the authors borrow from UOut's analysis and use to explain why placement and noise magnitude matter.

What would settle it

Rerun the best configuration (channel dropout on all blocks with ScheduledDropPath) and the no-dropout baseline on the same 10% class-balanced split for at least 30 seeds, and compare the distributions of best-epoch validation mIoU; if the difference in means is not significantly above zero, the claimed 0.10 improvement is not established.

Watch

Extended reading notes

Core claim

The paper's central discovery is that the effectiveness of dropout in a BatchNorm-heavy fully convolutional segmentation network is governed by where the dropout acts and how much variance it injects. Pixel-wise vanilla dropout fails because spatially correlated activations make it a learning-rate change rather than a regularizer; dropping whole channels removes entire features and forces the upper layers to recombine the remaining channels. In the authors' experiments, the best configuration -- channel dropout applied at the ResNet backbone, the spatial pyramid pooling, and the decoder, with dropout probability ramped linearly from 0 to 0.2 over 30 epochs -- improved validation mIoU from 0.49 to 0.59 over the unregularized model on 10% of the training data. They also report that DropBlock, which drops contiguous patches, injects noise that can hurt when applied broadly, and that UOut, designed to avoid variance shift after BatchNorm, gives smaller or inconsistent gains. The conclusion is that dropout remains useful at low data volumes, but its benefits depend on placement, scheduling, and the noise scale of the method.

Load-bearing premise

The paper's headline improvement is close to the run-to-run spread it reports, and it never states how many training runs each mean mIoU is averaged over, so the 0.49-to-0.59 gain may be seed noise rather than a real effect.

Editorial extensions

If this is right

  • In low-data segmentation, adding channel dropout to the encoder, pooling, and decoder with a 30-epoch ramp yields a validation mIoU gain of about 0.10 over no dropout, without changing the architecture or the data.
  • BatchNorm does not make dropout redundant; the interaction is negative mainly when dropout sits before BatchNorm and creates a train/test variance shift, which channel-wise placement and scheduling can mitigate.
  • A high-noise dropout method such as DropBlock can hurt mIoU when applied broadly, so choosing a low-variance method matters more than simply adding more regularization.
  • ScheduledDropPath is most beneficial for dropout methods that disrupt the image signal itself, such as DropBlock, because it delays regularization until after initial fitting.
  • The results imply that practitioners should not discard dropout purely because a model uses BatchNorm; channel-wise dropout with a ramp schedule is a viable low-data regularization even in such architectures.

Reading between the lines

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

  • If the 0.10 gain survives a properly seeded rerun, it implies that regularization placement is a first-order design choice in low-data dense prediction: channel dropout acts like a structural prior on feature independence, and the best location may be where the network learns class identity rather than fine boundaries.
  • The same recipe could transfer to other fully convolutional dense-prediction tasks with expensive labels, such as depth estimation, instance segmentation, or medical image segmentation, where low-data regimes are common; that is a testable extension rather than a result of this paper.
  • A natural diagnostic use follows from the paper's qualitative results: applying channel dropout at different stages shifts whether the model errs on classification or on edges, so per-stage ablation of dropout placement could reveal which layers encode category identity versus boundary detail.
  • One unresolved implication is that the paper compares scheduled versus unscheduled dropout but does not tune the ramp length or the dropout probability; those are obvious next experiments.
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 reports an empirical study of dropout variants (vanilla, SpatialDropout/channel dropout, DropBlock, UOut) applied at different positions of a DeepLabv3+ segmentation network under low-data conditions (10% of each PASCAL VOC 2012 class). The experiments cover a matrix of 16 configurations, with and without a linear dropout schedule (ScheduledDropPath). The headline claim is that scheduled channel dropout applied across all stages raises validation mIoU from 0.49 to 0.59 compared with the no-dropout baseline.

Significance. If the reported effect is real, the paper offers a useful practical result for low-data semantic segmentation and a systematic comparison of dropout placement that is more thorough than most prior work. The experimental matrix is a strength: all configurations are tabulated with mean, standard deviation, worst, median, and best mIoU, and qualitative epoch-by-epoch visualizations are provided for each configuration. However, the central quantitative claim is not yet statistically established. The reported standard deviations (≈0.09–0.11) are of the same magnitude as the headline improvement (0.10), the number of runs behind each mean is not stated, and the results are selected from many configurations after validation-based 'best epoch' selection. As reported, the main conclusion is indistinguishable from selection noise.

major comments (4)
  1. [Section 6.1, Tables 3 and 4] The paper reports mean and standard deviation of mIoU for each configuration but never states the number of runs (seeds). The headline gain of 0.10 (none 0.49 vs all-chandrop 0.59) is comparable to the reported standard deviations of 0.09–0.11. Without run counts, confidence intervals, or a paired significance test, the central empirical claim is not supported. Please provide the number of seeds, the individual run-level results, and a statistical test (e.g., paired t-test or Wilcoxon test) comparing the best configuration against the baseline.
  2. [Table 3 and Table 4 captions] The captions state 'Statistics were taken from the best epoch.' This means the validation set was used to select the epoch for each run, so the reported means are not estimates of the performance of a fixed model. This selection inflates the reported numbers relative to a standard evaluation protocol. I recommend either evaluating at a predetermined epoch (e.g., the final epoch) or reporting the full epoch-by-epoch curves and using a separate validation split for early stopping.
  3. [Section 6.1, Tables 3 and 4] The paper compares 16 dropout configurations and highlights the largest mean (all-chandrop with scheduling, 0.59) as the headline result. This is a best-of-many comparison. Under the null hypothesis that dropout has no effect, the maximum of 16 noisy estimates will be biased upward, so the observed ranking is partly a selection artifact. Please report multiple-comparison-corrected p-values or clearly state that the result is an exploratory finding that requires independent confirmation.
  4. [Section 6.1, paragraph 2] The text claims that 'application of dropout at the later layers only improves model generalization marginally compared to application of dropout at the earlier layers.' This is not directly supported by the means in Tables 3 and 4: in Table 3, resnet-chandrop (0.56) exceeds upper-chandrop (0.49) and decoder-chandrop (0.50) by a large margin, and in Table 4, decoder-chandrop (0.47) is below resnet-chandrop (0.53). Please either present a statistical comparison supporting this claim or temper the wording.
minor comments (5)
  1. [Section 3.4, Eq. for variance shift] The variance-shift formula is garbled: the printed expression 'v(∑w2i+ρx∑i∑j⁄=iwjwj(1p(µ2+v)−µ2)(∑w2i+ρax∑i∑j⁄=iwiwj)' is missing the quotient operator and is not readable. Please rewrite the equation using LaTeX with clear parentheses and a division sign.
  2. [Table 2, decoder-dropblock row] In Table 2, the row for decoder-dropblock lists 'Channel' in the ResNet Blocks column; it should be 'None' because this configuration applies DropBlock only to the decoder output.
  3. [Section 3.4, UOut implementation] The paper states that its UOut implementation draws one scalar from U[-p,p] per channel and applies it to the whole channel. Li et al.'s UOut as described in the same section adds per-element noise. Please clarify whether the channel-wise implementation is intentional and how it relates to the original UOut formulation.
  4. [Section 6.2.16] The sentence 'Applying UOut over all layers exhibits similar problems as seen in 6.2.15 and 6.2.15' repeats the same subsection number; one of the references should be to a different subsection (e.g., 6.2.14).
  5. [Throughout] There are several typos: 'Univeristy' in the author affiliations, 'simpifies' in Section 3.4, 'occurr' in Section 6.2.10, and 'foregrund' in Section 6.2.15. A careful proofread is needed.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity; the paper's claims are empirical measurements compared against external baselines, not derivations from their own outputs.

full rationale

The paper is a purely experimental study: the abstract's central claim (baseline mIoU 0.49 to 0.59 with scheduled SpatialDropout) is a reported validation-set measurement, not the output of a derivation. The only mathematical content in the paper, the variance-shift analysis in Section 3.4, is explicitly attributed to Li et al. [5] and is used only as background motivation for why dropout before BatchNorm can be harmful; it is not used to compute or predict the mIoU values in Tables 3 and 4. No parameter is fitted to the reported results and then renamed a prediction, no uniqueness theorem is invoked, and no self-citation carries an argument. The dropout methods (SpatialDropout, DropBlock, UOut) and ScheduledDropPath are adopted from external prior work [2, 3, 5, 12], and the DeepLabv3+ architecture is likewise cited externally [6]. The comparison across 16 configurations, with the best mean reported as the headline, raises a legitimate statistical-selection concern, but selecting the largest observed mean is not equivalent to defining the result from the result. The paper's derivation chain, such as it is, is self-contained: each experimental outcome is an independent measurement under stated conditions. Therefore no circular step is present.

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

The empirical claim depends on the specific training setup: a global dropout probability, a schedule length, learning rate multipliers, and a 10% class-balanced split. All are fixed by hand; the paper does not explore sensitivity to them. The only external inputs are the PASCAL VOC dataset and ImageNet-pretrained weights.

free parameters (4)
  • Dropout probability p = 0.2
    A single dropout probability was used for all methods and layers; the paper does not report tuning or sensitivity to p (Section 5.1.1).
  • ScheduledDropPath ramp length n = 30 epochs
    The schedule reaches full dropout strength at epoch 30, a hand-chosen hyperparameter (Section 3.5).
  • Learning rate and multipliers = 7e-3 base, 10x for SPP/decoder
    Training setup from the canonical PyTorch reimplementation, not justified experimentally (Section 5.1).
  • Per-class training fraction = 10%
    Defines the low-data regime that the claim depends on; no experiments with other fractions are reported (Section 5).
assumptions (4)
  • standard math Variance-shift analysis of dropout before BatchNorm (Li et al.) is correct and applicable here.
    Section 3.4 relies on this to motivate UOut; the equations in the text appear garbled but the citation is to a published paper.
  • domain assumption ImageNet-pretrained ResNet features transfer to PASCAL VOC segmentation.
    Section 5.1 states that without pretraining the model reaches at most 0.2 mIoU; the entire experiment depends on this transfer.
  • domain assumption PASCAL VOC 2012 labels and the validation split are a reliable measure of segmentation quality.
    All conclusions use validation mIoU as the target metric (Section 6.1).
  • ad hoc to paper The paper's UOut implementation, one scalar per channel drawn from U[-p,p], matches the intent of Li et al.
    Section 3.4 describes this specific implementation choice; it is not validated against the original authors' code.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Don't ignore Dropout in Fully Convolutional Networks." pith.science (2026). https://pith.science/paper/XTVBBQRM

@misc{pith2026190809162,
  author       = {Pith},
  title        = {Pith review of: Don't ignore Dropout in Fully Convolutional Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XTVBBQRM}},
  note         = {Machine review of arXiv:1908.09162}
}
read the original abstract

Data for Image segmentation models can be costly to obtain due to the precision required by human annotators. We run a series of experiments showing the effect of different kinds of Dropout training on the DeepLabv3+ Image segmentation model when trained using a small dataset. We find that when appropriate forms of Dropout are applied in the right place in the model architecture that non-insignificant improvement in Mean Intersection over Union (mIoU) score can be observed. In our best case, we find that applying Dropout scheduling in conjunction with SpatialDropout improves baseline mIoU from 0.49 to 0.59. This result shows that even where a model architecture makes extensive use of Batch Normalization, Dropout can still be an effective way of improving performance in low data situations.

Figures

Figures reproduced from arXiv: 1908.09162 by the authors.

Figure 2
Figure 2. Segmentation results over each epoch (pretrained-base). the later layers only improves model generalization marginally compared to application of dropout at the earlier layers. Third, using a form of dropout that does not increase noise variance that much is better. Introducing too much variance through the use of DropBlock in the earlier layers causes shifts in classification as we will see later on, leading to a d… view at source ↗
Figure 3
Figure 3. Best and worst segmentation results over each epoch ( [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 5
Figure 5. Segmentation results over each epoch (none). 6.2.2 Experiment none ( [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figures from the paper (15 more)
Figure 6
Figure 6. Figure 6: Best and worst segmentation results over each epoch ( [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 8
Figure 8. Figure 8: Segmentation results over each epoch (resnet-chandrop). tvmonitor class, as long as they are "assisted" by the presence of sharp gradients on the bezel of the monitor. In contrast to the none experiment, the model correctly is able to determine more pixels as belonging…
Figure 10
Figure 10. Figure 10: Segmentation results over each epoch (spp-chandrop ( [PITH_FULL_IMAGE:figures/full_fig_p010_10.png]
Figure 12
Figure 12. Figure 12: Segmentation results over each epoch (decoder-chandrop). indicating the Channel Dropout was able to improve the edge detection process but only for certain classes such as bus and motorbike and not cow. 6.2.6 Experiment upper-chandrop ( [PITH_FULL_IMAGE:figures/full_…
Figure 14
Figure 14. Figure 14: Segmentation results over each epoch (upper-chandrop). 10 [PITH_FULL_IMAGE:figures/full_fig_p010_14.png]
Figure 16
Figure 16. Figure 16: Segmentation results over each epoch (all-chandrop). (a) Segmentation for Image 1 (b) Segmentation for Image 2 [PITH_FULL_IMAGE:figures/full_fig_p011_16.png]
Figure 18
Figure 18. Figure 18: Segmentation results over each epoch (resnet-dropblock). 6.2.8 Experiment resnet-dropblock ( [PITH_FULL_IMAGE:figures/full_fig_p011_18.png]
Figure 20
Figure 20. Figure 20: Segmentation results over each epoch (spp-dropblock). 11 [PITH_FULL_IMAGE:figures/full_fig_p011_20.png]
Figure 22
Figure 22. Figure 22: Segmentation results over each epoch (decoder-dropblock). (a) Segmentation for Image 1 (b) Segmentation for Image 2 [PITH_FULL_IMAGE:figures/full_fig_p012_22.png]
Figure 24
Figure 24. Figure 24: Segmentation results over each epoch (all-dropblock). 6.2.10 Experiment decoder-dropblock ( [PITH_FULL_IMAGE:figures/full_fig_p012_24.png]
Figure 26
Figure 26. Figure 26: Segmentation results over each epoch (resnet-uout). 12 [PITH_FULL_IMAGE:figures/full_fig_p012_26.png]
Figure 28
Figure 28. Figure 28: Segmentation results over each epoch (spp-uout). (a) Segmentation for Image 1 (b) Segmentation for Image 2 [PITH_FULL_IMAGE:figures/full_fig_p013_28.png]
Figure 30
Figure 30. Figure 30: Segmentation results over each epoch (decoder-uout). 6.2.13 Experiment spp-uout ( [PITH_FULL_IMAGE:figures/full_fig_p013_30.png]
Figure 32
Figure 32. Figure 32: Segmentation results over each epoch (upper-uout). 13 [PITH_FULL_IMAGE:figures/full_fig_p013_32.png]
Figure 34
Figure 34. Figure 34: Segmentation results over each epoch (all-uout). We also see decent forground detection performance on Image 2, though the foreground object is again misclassified. 6.2.16 Experiment all-uout ( [PITH_FULL_IMAGE:figures/full_fig_p014_34.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

15 extracted references · 6 canonical work pages

  1. [1]

    Hinton, Oriol Vinyals, and Jeffrey Dean

    Geoffrey E. Hinton, Oriol Vinyals, and Jeffrey Dean. Distilling the knowledge in a neural network. CoRR, abs/1503.02531, 2015

  2. [2]

    Golnaz Ghiasi, Tsung-Yi Lin, and Quoc V . Le. Dropblock: A regularization method for convolutional networks. In Samy Bengio, Hanna M. Wallach, Hugo Larochelle, Kristen Grauman, Nicolò Cesa-Bianchi, and Roman Garnett, editors, NeurIPS, pages 10750–10760, 2018

  3. [3]

    Efficient object localization using convolutional networks

    Jonathan Tompson, Ross Goroshin, Arjun Jain, Yann LeCun, and Christoph Bregler. Efficient object localization using convolutional networks. In CVPR, pages 648–656. IEEE Computer Society, 2015

  4. [4]

    Batch normalization: Accelerating deep network training by reducing internal covariate shift

    Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In Francis R. Bach and David M. Blei, editors, ICML, volume 37 of JMLR Workshop and Conference Proceedings, pages 448–456. JMLR.org, 2015. 14

  5. [5]

    Understanding the disharmony between dropout and batch normalization by variance shift

    Xiang Li, Shuo Chen, Xiaolin Hu, and Jian Yang. Understanding the disharmony between dropout and batch normalization by variance shift. CoRR, abs/1801.05134, 2018

  6. [6]

    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. CoRR, abs/1802.02611, 2018

  7. [7]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. CoRR, abs/1512.03385, 2015

  8. [8]

    Xception: Deep learning with depthwise separable convolutions

    François Chollet. Xception: Deep learning with depthwise separable convolutions. In Proceed- ings of the IEEE conference on computer vision and pattern recognition , pages 1251–1258, 2017

Show all 15 references
  1. [9]

    Spatial pyramid pooling in deep convolutional networks for visual recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Spatial pyramid pooling in deep convolutional networks for visual recognition. CoRR, abs/1406.4729, 2014

  2. [10]

    Dropout: a simple way to prevent neural networks from overfitting

    Nitish Srivastava, Geoffrey E Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdi- nov. Dropout: a simple way to prevent neural networks from overfitting. Journal of Machine Learning Research, 15(1):1929–1958, 2014

  3. [11]

    Making convolutional networks shift-invariant again

    Richard Zhang. Making convolutional networks shift-invariant again. In ICML, 2019

  4. [12]

    Barret Zoph, Vijay Vasudevan, Jonathon Shlens, and Quoc V . Le. Learning transferable architectures for scalable image recognition. In CVPR, pages 8697–8710. IEEE Computer Society, 2018

  5. [13]

    Everingham, L

    M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2012 (VOC2012) Results. http://www.pascal- network.org/challenges/VOC/voc2012/workshop/index.html, 2012

  6. [14]

    Deeplab v3+ in pytorch, 2018

    Jianfeng Zhang. Deeplab v3+ in pytorch, 2018

  7. [15]

    Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. CoRR, abs/1606.00915, 2016. 15

Pith tools

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