acceptodds
Under review as a conference paper at ICLR 2027

PixCalibrate: Pixel-Mixing Guided Calibration And Robustness for Image Corruption

Abstract

Reliable image classification requires models to remain accurate and calibrated under distribution shift. However, models trained on clean data are often prone to incorrect and overconfident predictions on corrupted or out-of-distribution test data. Existing augmentation frameworks partially address this issue, but they typically rely on global interpolation or introduce sharp region-level changes, neither of which simulates real-world corruptions. In this work, we propose PixCalibrate, a training-time pixel-level mixing approach for improving both robustness and calibration under corrupted settings. PixCalibrate generates augmented samples that preserve the semantic structure of an anchor image, while injecting a controlled fraction of pixels from a perturbation source—auxiliary image, producing fine-grained, spatially distributed perturbations that closely mimic real-world corruptions without disrupting the overall image structure. PixCalibrate consistently achieves strong accuracy and calibration across standard image corruption benchmarks and outperforms multiple prior augmentation frameworks.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Image Corruption, Calibration, Mixup, Data Augmentation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_CcY1OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The manuscript proposes PixCalibrate, a training-time pixel-level mixing augmentation method. Its core idea is to use AUM-guided easy-anchor/hard-source pixel-level mixing to simulate fine-grained, spatially distributed corruptions, and to combine it with a confidence penalty to simultaneously improve accuracy and calibration under corruption. The labels are mixed in proportion to the replacement fraction. The authors evaluate accuracy and calibration on several corruption benchmarks, additional shifted test sets, and ImageNet variants. The experiments are reasonably extensive, but the novelty looks limited: the method is essentially a combination of AUM, pixel mixing, and an entropy penalty.

Critical strengths
  • The method is simple and relatively inexpensive.
  • The authors report corruption-wise results and an analysis excluding pixel-level corruption types.
  • The experimental coverage is reasonably extensive, with evaluation on CIFAR-10.1, CIFAR-10.2,CINIC-10, and CIFAR-10-W.
  • Tccuracy is highest among the compared-to methods in each case, albeit usually by a small margin? That said, the comparison is restricted to Mixup and CutMix.
Critical weaknesses
  • The method is mainly a combination of AUM, pixel-level mixing, and a confidence penalty, and is closely related to CalibrateMix, Mixup, CutMix, etc.; its theoretical novelty is limited.
  • The sampling procedure of the mask is not sufficiently specified.
  • AUM splits easy/hard samples by the median within each batch. However, when the batch is small, classes are imbalanced, or label noise is high, the split may be unstable.
  • The comparison with CalibrateMix may not be entirely fair, since the original CalibrateMix setting is semi-supervised, whereas this paper is fully supervised.
  • The mechanism is not isolated. Replacing 40% of randomly located pixels with values from another image creates a noise-like perturbation. It seems that the largest CIFAR-10-C improvements occur for Gaussian, shot, and impulse noise, whereas brightness barely changes. The same-class and different-class versions are also close in accuracy. This is consistent with noise-like augmentation being important, although it does not prove that label mixing is unnecessary. A matched pixel-noise replacement control, with and without mixed labels, might be essential to distinguish the proposed mechanism from a simpler augmentation.
  • The advantage of AUM-based pairing is not convincingly established. The CIFAR-10-C ECE improvement is worth noting, but most differences are small. Without seed-level paired uncertainty for this ablation, the evidence does not yet support a strong claim that pairing is a key ingredient.
  • Relevant robustness and calibration baselines deserve more attention. The results on Table 55 do not seem to support a general claim of state-of-the-art calibration is not supported on this benchmark. AugMix and PixMix should be included in the main comparison under a clearly matched training recipe, alongside a simple noise baseline; temperature scaling and RegMixup would also be informative. Some existing baselines perform below the unaugmented model, which warrants an explanation of tuning and implementation. Worth mentioning is that published numbers from different WRN variants are not directly comparable.
  • The ImageNet evidence is limited by the reference training recipe. The reported clean ResNet-50 baseline is 71.76%, and its ImageNet-C accuracy is 26.93%. These are weak relative to commonly used ResNet-50 training results, making it difficult to judge how much of the improvement would persist against stronger baselines. The ImageNet comparison also lacks AugMix/PixMix and seed variability.
  • The statistical analysis requires more care. Testing across corruption–severity cells from the same trained models does not provide independent training replications. Seed-level results, ideally with paired model comparisons, would be a more appropriate basis for significance claims. The similarity analysis also lacks enough detail about corruption conditions and variability to interpret the very small frequency-distance differences.
Questions for authors
  • What is the specific sampling distribution of the binary mask? Is it an independent per-pixel Bernoulli?
  • Have you tried spatially correlated masks to better simulate corruptions with spatial structure, such as fog, weather, and blur?
  • What proportion of the hard samples are mislabeled/ambiguous?
  • How does the method compare with random pixel replacement, impulse noise, and same-image pixel shuffling at the same replacement fraction?
  • Can the authors report seed-level uncertainty for the AUM-versus-random-pairing comparison?
  • How do AugMix, PixMix, and RegMixup perform under exactly the same CIFAR and ImageNet training settings?
  • Why is the ResNet-50 clean baseline at 71.76%, and what changes under a stronger conventional recipe?
5 comments
AuthorsOpenReview ↗

Author response to Reviewer CcY1

We thank the reviewer for the valuable feedback. We address the weaknesses and questions below and we will incorporate them in the paper.

>>W1:

We thank the reviewer for raising this point. PixCalibrate methodological contribution centers on the design of a new augmentation technique, which naturally resembles test-time corrupted images. While CalibrateMix and random Mixup directly interpolate pixel values across two images, i.e., all contiguous pixels of an image are overlaid with all contiguous pixels from another image, according to the mixing ratio (see Fig. 2), PixCalibrate samples (uniformly at random) individual pixels from a hard image and places them into an anchor image (by completely replacing the anchor’s pixels with the pixels sampled from the hard image in exact same positions). More precisely, PixCalibrate uniformly samples individual pixel locations and replaces their values with those from a paired image, producing spatially distributed perturbations while retaining much of the anchor's structure (see also Fig. 2), especially because we preserve 60% of the anchor's pixels intact and replace only 40% with the pixels sampled from the hard image. Please note that, unlike CutMix, these 40% replacements are spatially scattered rather than concentrated in contiguous patches that create an entire patch of pixels being replaced. Figure 2 visually illustrates this distinction, showing that PixCalibrate introduces localized variations while retaining the anchor's overall appearance. Table 3 provides quantitative support: PixCalibrate achieves higher structural similarity to corruption perturbations (SSIM: 0.36) than Mixup (0.20) and CutMix (0.27).

In addition, we incorporate difficulty-aware pairing, which is inspired from CalibrateMix. However, while CalibrateMix combines easy labeled samples with hard pseudo-labeled samples, and hard labeled samples with easy pseudo-labeled samples, PixCalibrate uses a fixed direction to the construction: the easy sample provides the anchor, while the hard sample provides the perturbation. This design aims to preserve the features of the anchor while introducing challenging variations from the hard sample. As demonstrated in Table 5, reversing this direction by using hard samples as anchors can produce more ambiguous augmented samples, reducing label reliability. At our default replacement ratio (ρ=0.4\rho=0.4), 60% of spatial locations retain the easy image’s original pixel values, while the remaining locations receive pixels from the hard image. Thus, difficulty determines both which samples are paired, which sample acts as anchor, and which sample acts as the perturbation source.

Importantly, these methodological differences contribute to improvements beyond those provided by confidence regularization. Although the confidence penalty complements our pixel-sampling strategy by encouraging higher predictive entropy, removing it still yields substantial improvements over CalibrateMix (as shown in the table below and Tables 1 and 5 in the paper):

MethodCifar-10c accCifar-10c eceCifar-100c accCifar-100c ece
CalibrateMix75.5411.7751.6018.01
PixCalibrate without confidence penalty83.238.4363.0110.15
PixCalibrate84.707.6465.038.88

Furthermore, Appendix A.6.2 applies the same confidence penalty to Mixup and CutMix; these variants still underperform PixCalibrate without the penalty on both accuracy and ECE. This supports treating confidence regularization as a complementary component rather than the primary explanation for the gains.

We will revise the presentation to distinguish the inherited ingredients from the proposed construction. Specifically, our contribution is the integration of difficulty-dependent anchor/perturbation roles with spatially distributed mixing for jointly improving robustness and calibration under corruption. The pixel-sampling and replacement operator is central to this contribution, and the role-selection ablation provides evidence for the additional design choice beyond mixing granularity.

AuthorsOpenReview ↗

Author response to Reviewer CcY1

>>W2, Q1: binary mask
We thank the reviewer for pointing out the need for clarification. As described in Eq. 4, PixCalibrate generates a binary mask M∈0,1H×WM ∈ {0, 1}^{H×W} by uniformly sampling spatial pixel locations, with a fixed replacement fraction (ρ=0.4\rho=0.4). The same mask is applied across all RGB channels, ensuring that complete pixel values are replaced rather than individual color channels.

>>W3: unstable split on noisy and small data batch

We thank the reviewer for raising this concern. We acknowledge that batch-wise median splitting may introduce variability in easy/hard assignments. However, despite these potential variations, our experimental results demonstrate that the simple median-based difficulty split consistently outperforms existing baselines across the evaluated corruption benchmarks in both accuracy and calibration.

>>W4: CalibrateMix on fully supervised

We thank the reviewer for pointing out the need for clarification. We preserve CalibrateMix's original AUM-based difficulty estimation, top-k feature-dissimilarity selection, bidirectional pairing strategy, and global Mixup operation. Since all training samples are labeled, the original easy-labeled–hard-unlabeled and easy-unlabeled–hard-labeled pairings are adapted to easy-labeled–hard-labeled pairings in both directions, using ground-truth labels instead of pseudo-labels. The SSL-specific unsupervised objectives are omitted, while the remaining mixing loss is retained. All methods follow the same fully supervised training protocol. We will clarify these adaptation details in the revised manuscript.

>>W5, Q4: Random Noise instead of pixels

We thank the reviewer for this observation. Unlike Gaussian or salt-and-pepper noise, which introduce perturbations from predefined distributions, PixCalibrate generates image-dependent perturbations using pixels from another training sample. As illustrated in Fig. 2, this produces spatially distributed variations while retaining recognizable anchor content.

To further investigate whether the improvements can be attributed to simpler noise-like augmentations or label mixing, we conducted additional experiments on CIFAR-10-C and report the results below:

MethodCifar10c eceCifar10c acc
FullySupervised15.1873.48
FullySupervised + gaussian noise augmentation11.2576.95
Pixel shuffle8.9383.52
Random pixel values8.6683.16
Original label7.8183.35
PixCalibrate (ours)7.6484.70

The results show that Gaussian noise augmentation improves upon the FullySupervised baseline, confirming that noise-like perturbations contribute to corruption robustness. Pixel shuffling and random pixel replacement provide further improvements, but both remain behind PixCalibrate in accuracy and calibration. Furthermore, using only the original anchor label while keeping pixel-sampling and replacement unchanged reduces accuracy from 84.70% to 83.35% and increases ECE from 7.64 to 7.81. These findings suggest that image-dependent pixel replacement is beneficial beyond simpler perturbations, while proportional soft-label mixing provides an additional improvement.

We acknowledge that the largest improvements occur on noise-like corruptions. However, the modest gain on brightness is partly explained by the already high baseline accuracy (94.13%), which PixCalibrate maintains at 94.22%. Importantly, the benefits extend beyond noise-like corruptions. Appendix A.6.1 shows that, after excluding five pixel-level corruption types, PixCalibrate achieves 86.80% accuracy and 5.44 ECE. These results support the broader effectiveness of PixCalibrate beyond noise-like perturbations.

>>W6: Benefit of AUM-pairing over random pairing

We thank the reviewer for this observation. We would like to clarify that AUM-guided pairing is a complementary component rather than the primary novelty of PixCalibrate, which lies in its pixel-level sampling and replacement mechanism. Nevertheless, AUM-guided pairing consistently improves calibration across all three seeds on CIFAR-10-C, reducing ECE from 8.48 ± 0.29 (random pairing) to 7.63 ± 0.73

AuthorsOpenReview ↗

Author response to Reviewer CcY1

>>W7, Q6: PixMix, AugMix, and temperature scaling

We thank the reviewer for this suggestion. First, we clarify that all corresponding experiments have been performed under the same training settings. Second, AugMix and PixMix are already included in Table 5 under the same training settings as PixCalibrate. We present them separately to distinguish robustness-oriented augmentation strategies from the conventional image-mixing baselines in Table 1. We will make this distinction and the reference to Table 5 more explicit in the revised manuscript.

To further strengthen the evaluation, we conducted additional experiments with post-hoc temperature scaling on CIFAR-10-C. For temperature scaling, we optimize a single temperature parameter for each trained model by minimizing negative log-likelihood on a validation set. The results are reported in the table below:

MethodCifar10c eceCifar10c acc
FullySupervised15.1873.48
FullySupervised + Temp Scaling6.7073.48
PixCalibrate + Temp Scaling4.8984.70
PixCalibrate7.6484.70

The results show that, temperature scaling reduces the FullySupervised baseline's ECE from 15.18 to 6.70.In contrast, PixCalibrate achieves 84.70% accuracy, and applying temperature scaling further reduces its ECE from 7.64 to 4.89, outperforming the temperature-scaled FullySupervised baseline in both accuracy and calibration. These findings demonstrate that PixCalibrate's robustness improvements are complementary to post-hoc confidence calibration. We will update the manuscript with the added experiments.

>>W8, Q7: Imagenet training

We thank the reviewer for this observation. We clarify that our ImageNet experiments use the standard ResNet-50 architecture and follow the training protocol used in CalibrateMix on the FullySupervised setting on ImageNet.

Importantly, all methods use identical architectures, optimization settings, learning-rate schedules, and training durations, ensuring a controlled comparison. Our evaluation focuses on the relative improvements under matched training conditions. To further strengthen the ImageNet evaluation, we conducted additional experiments with CutMix and ResizeMix under the same training configuration and report the results below:

Imagenet eceImagenet accImagenet-c eceImagenet-c acc
CutMix5.4071.8922.6728.92
ResizeMix5.1972.1521.1829.61
PixCalibrate (ours)3.7472.7915.8632.53

As can be seen from the table above, PixCalibrate consistently outperforms both augmentation baselines in accuracy and calibration on clean and corrupted ImageNet. Compared with the stronger baseline, ResizeMix, PixCalibrate improves ImageNet-C accuracy by 2.92 points while reducing ECE by 5.32 points. These findings provide additional evidence that the benefits of PixCalibrate extend to large-scale image classification under matched training conditions.

>>W9: Statistical Significance and Frequency Distance.

We thank the reviewer for this observation. As per your suggestion, we performed statistical analysis using paired comparisons across three independent training seeds. The revised tests confirm that PixCalibrate's improvements over the FullySupervised baseline remain statistically significant (pp<0.05)), supporting the reliability of our reported accuracy and calibration gains. We will update the manuscript to reflect this corrected procedure.

Regarding the similarity analysis, we agree that the frequency-distance difference between PixCalibrate (0.1716) and Mixup (0.1717) is marginal and does not provide strong evidence of superiority. However, the more notable result is PixCalibrate's higher SSIM (0.36), compared with Mixup (0.20) and CutMix (0.27) which provides evidence that PixCalibrate's augmentation perturbations exhibit greater structural similarity to the evaluated corruption perturbations.

>> Q2:

We thank the reviewer for the question. PixCalibrate uses randomly distributed pixel-level masks rather than spatially correlated masks, as our objective is to generate diverse, fine-grained perturbations without explicitly modeling specific corruption types. Nevertheless, our results show improvements on spatially structured corruptions, including defocus blur (83.44% to 87.56%), zoom blur (77.53% to 84.15%), snow (82.28% to 88.86%), and fog (86.57% to 89.25%) on CIFAR-10-C. Furthermore, Appendix A.6.1 demonstrates that the improvements persist after excluding five pixel-level corruption types. These findings suggest that independently sampled pixel masks can also improve robustness to broader corruption patterns.

Reviewer_x6RiOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

The paper proposes PixCalibrate, a training-time augmentation for improving accuracy and calibration under image corruptions. It replaces a random, spatially scattered subset of pixels in an “easy” anchor image with pixels from a “hard” image, where difficulty is estimated by Area Under the Margin (AUM), and trains on the mixed images with area-proportional soft labels and a confidence penalty. The authors argue that such pixel-level mixing mimics real-world corruptions better than global interpolation (Mixup) or patch-based mixing (CutMix) while preserving the anchor’s semantics. Experiments on CIFAR-10/100-C, newly constructed SVHN-C/STL-10-C, ImageNet-C/A/O and several natural-shift test sets report improved accuracy and ECE over mixing-based baselines, supported by ablations on the pairing strategy, confidence penalty and hyperparameters.

Critical strengths
  1. Simple, cheap, and consistently effective on noise-type corruptions. The method requires only a random mask and per-sample margin tracking, with no external data or auxiliary networks. Gains on noise and other high-frequency corruptions are large and consistent across four datasets and 3 seeds (e.g., CIFAR-100-C Gaussian noise 23.03 → 49.69, glass blur 22.31 → 59.49; Table 8). Unlike pure Gaussian-noise augmentation, which is known to degrade fog and contrast (Yin et al., 2019), it does not appear to hurt these corruptions on CIFAR-10-C. This is potentially interesting but is not analyzed.

  2. Transparent reporting that enables analysis. Per-corruption accuracy and ECE for all methods (Tables 7–10), a confidence-penalty control applied to baselines (Table 12) and a natural-shift evaluation (Table 4) allow readers to verify where the gains come from. Table 12 also shows that the improvement is not explained by the confidence penalty alone. These strengths are consequential because they establish that the effect is real. As argued below, however, they also show that its cause differs from the one the paper claims.

Critical weaknesses
  1. The motivation for pixel-level mixing and its relation to existing techniques are unclear. I am not fully convinced by the premise that pixel-level mixing preserves the anchor’s semantics while mimicking real-world corruptions. Image semantics are largely carried by spatial relationships among neighboring pixels. When 40% of pixels are replaced at random locations, little of the spatial structure of xhx_h seems to remain, and the anchor’s local structure is also affected in most regions. Moreover, in expectation over the mask, the mixed image equals Mixup with λ = 0.6, with zero-mean noise added in each realization: x~=0.6xa+0.4xh+(M−0.4)⊙(xh−xa)\tilde x = 0.6x_a + 0.4x_h + (M-0.4)\odot(x_h-x_a). It is therefore not clear to me how the method preserves the anchor’s semantics better than Mixup, as suggested in §1 and §3.2.

Relatedly, §3.3 describes xhx_h as contributing “a masked perturbation,” while Eq. 6 assigns 40% of the label to yhy_h. This soft target may act similarly to label smoothing, and it appears to be in some tension with the invariance objective in Eq. 3. The choice of ρ = 0.4 would also benefit from a justification beyond the search on the corrupted test sets (Table 14).

The per-corruption results suggest the method may behave similarly to noise augmentation. A large share of the gains (58–72% of the accuracy improvement) comes from Gaussian/shot/impulse noise, glass blur and pixelate (Table 8), while gains on fog, brightness, contrast and the natural-shift benchmarks (Table 4) are small. Given this, comparisons with noise augmentation methods such as Gaussian / Patch-Gaussian augmentation and Rusak et al. (2020) seem essential. Without them, it is difficult to assess the specific benefit of mixing pixels between images.

  1. Some claimed contributions and parts of the evaluation would benefit from stronger support. The difference between AUM-guided and random pairing in Table 5 appears to be within the seed standard deviations reported in Table 6, and warmup settings other than the default perform comparably to or below random pairing (Table 13). This makes it hard to attribute the gains to AUM-guided pairing. In addition, ρ, β and warmup appear to be selected on the corrupted test sets, and the baseline hyperparameters are not reported. On ImageNet, only a FullySupervised baseline is included, and its accuracy seems lower than that of a standard ResNet-50 recipe, which makes the size of the improvement harder to interpret.

Update after rebuttal: My main concerns have been addressed. In the authors’ additional experiment, a baseline combining fixed-ratio (λ = 0.6) Mixup with Gaussian noise of matched per-pixel variance, under PixCalibrate’s otherwise identical pipeline, performed significantly worse (79.03 / 21.25 vs. 84.70 / 7.64 for PixCalibrate on CIFAR-10-C). This confirms that the proposed method is differentiated from simply combining Mixup and noise. An ablation study of anchor labels revealed that most of the performance improvement comes from input perturbation. The authors agreed to redefine AUM-guided pairing as complementary. Some limitations remain: hyperparameters were selected on the test split rather than a validation split, and there is no evaluation on held-out corruptions. Accordingly, the rating is raised from 2 to 3 and soundness from 1 to 2.

Questions for authors
  1. Role of the second image. §3.3 describes xhx_h as contributing a masked perturbation. How would the results change if the same masked pixels were filled with random noise (e.g., Gaussian or salt-and-pepper) instead of pixels from another image? A comparison with standard noise augmentation, such as Gaussian or Patch-Gaussian, would also help clarify the contribution.

  2. Relation to existing techniques. Since the mixed image can be written as fixed-ratio (0.6:0.4) Mixup plus zero-mean noise, would training with this combination directly (with random pairing) yield similar results? If PixCalibrate performs clearly better, that would strengthen the paper considerably.

  3. Soft label design. Could the authors explain the rationale for assigning 40% of the label to yhy_h when xhx_h is described as a perturbation? It would be helpful to see results using only the anchor label, using standard label smoothing, and with the pixel and label ratios set independently.

  4. Generalization to unseen corruptions. Do the gains remain on corruptions that do not closely resemble the augmentation, when hyperparameters are selected without the corrupted test sets?

  5. AUM and ImageNet. Is the difference between AUM-guided and random pairing statistically significant across seeds? Could the authors also report ImageNet results with a standard ResNet-50 recipe and additional augmentation baselines?

7 comments
AuthorsOpenReview ↗

Author response to Reviewer x6Ri

We thank the reviewer for the valuable feedback. We address the weaknesses and questions below and we will incorporate them in the paper.

>>W1: Similarity to Mixup and noise augmentation methods

We thank the reviewer for these observations. We would like to point out the differences between Mixup and PixCalibrate as follows. Mixup directly interpolates pixel values across two images, i.e., all contiguous pixels of an image are retained and overlaid with all contiguous pixels from another image, according to the mixing ratio (see Fig. 2). Specifically, a pixel from an image contributes to the newly formed image in proportion of λ\lambda and is added to its counterpart pixel from the other image which contributes in proportion of (1-λ\lambda). In contrast, PixCalibrate samples (uniformly at random) individual pixels from a hard image and places them into an anchor image (by completely replacing the anchor’s pixels with the pixels sampled from the hard image in exact same positions). More precisely, PixCalibrate uniformly samples individual pixel locations and replaces their values with those from a paired image, producing spatially distributed perturbations while retaining much of the anchor's structure (see also Fig. 2), especially because we preserve 60% of the anchor's pixels intact and replace only 40% with the pixels sampled from the hard image. Please note that, unlike CutMix, these 40% replacements are spatially scattered rather than concentrated in contiguous patches that create an entire patch of pixels being replaced. Figure 2 visually illustrates these distinctions, showing that PixCalibrate introduces localized variations while retaining the anchor's overall appearance. Table 3 provides quantitative support: PixCalibrate achieves higher structural similarity to corruption perturbations (SSIM: 0.36) than Mixup (0.20) and CutMix (0.27).

We would also like to point out that PixCalibrate introduces discrete, image-dependent perturbations whose values depend on the paired images, rather than independent additive noise. Please note that each output pixel in the combined image retains a value from one of the two original images instead of an interpolated value (we do not multiply pixel values with a mixing ratio, rather MM is a binary mask, M∈0,1H×WM \in {0,1}^{H \times W}. More precisely, in our case, all the pixels in a combined image are pixels from images within the dataset itself, whereas in Mixup, they are not, unless we make a strong assumption of convexity about the dataset that all points within the convex hull of the dataset are also valid datapoints. Moreover, because neural networks are nonlinear, training on these stochastic realizations is not generally equivalent to training on their expected image. We will clarify this in the paper. Thank you again for these comments.

>>W2. Contribution of AUM-Guided Pairing and Experimental Setup.

We thank the reviewer for the observations. We would like to clarify that AUM-guided pairing is a complementary component, while the primary methodological contribution of PixCalibrate lies in its pixel-sampling and replacement mechanism. Nevertheless, AUM-guided pairing improves calibration across all three seeds on CIFAR-10-C, reducing ECE from 8.48 ± 0.29 (random pairing) to 7.64 ± 0.73. Furthermore, Table 13 demonstrates that PixCalibrate maintains competitive performance across different warmup durations, suggesting that its effectiveness is not restricted to a single configuration.

Regarding the experimental setup, Appendix A.2 reports the architectures and optimization settings. All ImageNet experiments use the same ResNet-50 architecture and identical training configuration for both the FullySupervised baseline and PixCalibrate, ensuring that the observed improvements are attributable to the proposed method under the evaluated setup. Specifically, PixCalibrate improves ImageNet-C accuracy from 26.93% to 32.53%, while reducing ECE from 22.48 to 15.86.

AuthorsOpenReview ↗

Author response to Reviewer x6Ri

>> Q1

We thank the reviewer for this suggestion. Unlike Gaussian or salt-and-pepper noise, PixCalibrate introduces image-dependent perturbations, where replacement pixel values are determined by paired images rather than sampled from a predefined noise distribution. To further investigate the role of noise augmentation, we conducted additional experiments on CIFAR-10-C by replacing the same masked pixels with gaussian noise. The results are reported in the table below:

MethodCifar10c eceCifar10c acc
FullySupervised15.1873.48
FullySupervised + gaussian noise11.2576.95
PixCalibrate7.6484.70

The results from the table show that Gaussian noise augmentation improves upon the FullySupervised baseline, increasing accuracy from 73.48% to 76.95% and reducing ECE from 15.18 to 11.25. However, PixCalibrate achieves substantially greater improvements, reaching 84.70% accuracy and 7.64 ECE. This suggests that image-dependent pixel replacement provides additional advantages for both robustness and calibration.

Furthermore, our objective is to improve robustness without explicitly incorporating CIFAR-C corruption operators during training. As shown in Appendix A.6.1, even after excluding five pixel-level corruptions, PixCalibrate achieves 86.80% accuracy and 5.44 ECE, outperforming Mixup (83.76% / 6.32). These findings support the benefits of PixCalibrate beyond noise-like corruptions.

>> Q2

We thank the reviewer for this insightful suggestion. We compare PixCalibrate with combining fixed-ratio Mixup with zero-mean noise on CIFAR-10-C. The results are presented in the table below:

MethodCifar10c eceCifar10c acc
Mixup11.1575.41
Mixup with zero mean noise21.2579.03
PixCalibrate7.6484.70

The results show that adding zero-mean noise to Mixup improves accuracy from 75.41% to 79.03%, but substantially worsens calibration, increasing ECE from 11.15 to 21.25. In contrast, PixCalibrate achieves 84.70% accuracy and 7.64 ECE, outperforming Mixup with zero-mean noise by 5.67 percentage points in accuracy and reducing ECE by 13.61 points. These findings indicate that simply combining Mixup with zero mean noise does not reproduce the joint robustness and calibration improvements of PixCalibrate. Unlike independently generated noise, our pixel-level perturbations are derived from paired-image differences, introducing image-dependent variations while retaining the anchor's overall structure.

>> Q3

We thank the reviewer for this question. Although the hard sample serves as a perturbation source, its pixel values contribute to the augmented image and may introduce semantic ambiguity. Thus, the 40% label weight reflects the replacement ratio rather than an exact measure of semantic content. Importantly, unlike standard label smoothing, which distributes probability mass across classes, our approach assigns weights specifically to the two paired classes.

Per your suggestion, to evaluate this design choice, we conducted an additional experiment using only the original anchor label while keeping the pixel-sampling operation unchanged and we also experimented with standard label smoothing with the FullySupervised baseline. We report the results below:

MethodCifar10c eceCifar10c acc
FullySupervised15.1873.48
FullySupervised + Label Smoothing = 0.111.4274.07
Mixup11.1575.41
Original anchor label7.8183.35
PixCalibrate (ours)7.6484.70

The results show that standard label smoothing improves calibration over the FullySupervised baseline but provides only marginal accuracy gains. Furthermore, under identical pixel-sampling settings, proportional soft-label mixing improves accuracy by 1.35 points while also reducing ECE from 7.81 to 7.64. This suggests that accounting for the contribution of the second image through soft targets is beneficial, rather than treating the augmented sample as entirely belonging to the anchor class.

AuthorsOpenReview ↗

Author response to Reviewer x6Ri

>> Q4

We thank the reviewer for this question. The hyperparameters were not optimized for specific corruption types or severity levels or averaged corrupted test set performance. As per your suggestion, we report the corresponding clean test accuracy below:

HyperparametervalueCifar10 acc
ρ\rho0.0595.22
ρ\rho0.2095.38
ρ\rho0.4095.62
ρ\rho0.4595.41
ρ\rho0.5095.55
β\beta0.0195.47
β\beta0.195.62
β\beta195.11

As can be seen from the table, (ρ=0.4\rho=0.4, β=0.1\beta=0.1) achieve the highest clean accuracy (95.62%) among the evaluated configurations. Thus, perturbing samples with 40% pixels from the perturbation source provides the best performance on Cifar10.

We emphasize that PixCalibrate is trained exclusively on clean images, without explicitly introducing CIFAR-C corruption operators during training. To examine generalization beyond noise-like corruptions, Appendix A.6.1 evaluates performance after excluding five pixel-level corruptions. PixCalibrate achieves 86.80% accuracy and 5.44 ECE on the remaining corruptions, outperforming Mixup (83.76% / 6.32). Furthermore, Table 4 demonstrates improvements on four non-synthetic distribution shifts (CIFAR-10.1, CIFAR-10.2, CINIC-10, and CIFAR-10-W), indicating that the benefits extend beyond synthetic corruption benchmarks.

>> Q5

We clarify that AUM-guided pairing is a complementary component rather than the primary contribution of PixCalibrate. As observed, PixCalibrate achieves statistically significant improvements over the FullySupervised baseline on CIFAR-10-C, CIFAR-100-C, SVHN-C, STL-10-C. Per your suggestion, we are running experiments on ImageNet with multiple seeds to perform the statistical significance test of AUM-guided and random-pairing across seeds. We will report the results in the updated manuscript.

On ImageNet, both the FullySupervised baseline and PixCalibrate use the same ResNet-50 architecture and training configuration (details provided in Appendix A.2), ensuring a fair comparison.

As per your suggestion, under the same training setup, we compare PixCalibrate with additional augmentation baselines, CutMix and ResizeMix on both clean test set of ImageNet and corrupted test set, ImageNet-C. We report the results below:

Imagenet eceImagenet accImagenet-c eceImagenet-c acc
CutMix5.4071.8922.6728.92
ResizeMix5.1972.1521.1829.61
PixCalibrate3.7472.7915.8632.53

As can be seen from the table, PixCalibrate consistently outperforms both augmentation baselines in accuracy and calibration on clean and corrupted ImageNet. Compared with the stronger baseline, ResizeMix, PixCalibrate improves ImageNet-C accuracy by 2.92 points while reducing ECE by 5.32 points. These findings provide additional evidence that the benefits of PixCalibrate extend to large-scale image classification under matched training conditions.

Reviewer_nhYXOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PixCalibrate, a training-time data augmentation framework designed to improve classification robustness and confidence calibration under image corruptions. The method combines AUM-based difficulty-aware sample pairing, pixel-level mixing between easy and hard samples, and confidence regularization. Experiments on several corruption benchmarks demonstrate improvements in classification accuracy and calibration.

Critical strengths
  1. The paper addresses the important challenge of maintaining both predictive accuracy and reliable confidence under distribution shifts and image corruptions.
  2. The proposed framework is straightforward to implement and does not require additional models or external training data.
  3. PixCalibrate achieves substantial improvements on multiple corruption benchmarks, including CIFAR-10-C, CIFAR-100-C, and ImageNet-C. The paper also includes ablation studies and evaluations on non-synthetic distribution shifts.
Critical weaknesses
  1. The overall framework is closely related to CalibrateMix (AAAI 2026), which already introduces AUM-based difficulty-aware mixing for improving calibration. The main differences lie in the pixel-level mixing operation and additional confidence regularization, making the extent of the novel contribution somewhat unclear.

  2. Although CalibrateMix is included as a baseline, the experiments do not fully isolate the benefits of pixel-level mixing from differences in pairing strategies and confidence regularization. The adaptation of CalibrateMix to the fully supervised setting is also insufficiently explained.

  3. Randomly replacing pixels with content from another image does not necessarily reproduce the statistical properties of real-world corruptions. Moreover, assigning mixed labels according to the fraction of replaced pixels may not accurately reflect the semantic content of the augmented image.

Questions for authors
  1. Could the authors clarify the key methodological novelty of PixCalibrate relative to CalibrateMix, particularly beyond replacing global interpolation with pixel-level mixing?
  2. Could the authors provide a controlled comparison between global Mixup and pixel-level mixing under identical AUM-based pairing, training settings, and confidence regularization? How exactly was CalibrateMix adapted for the fully supervised experiments?
  3. What evidence supports the claim that cross-image pixel replacement effectively approximates real-world corruptions? Could the authors further justify the use of pixel-proportion-based soft labels and investigate alternative labeling strategies?
3 comments
AuthorsOpenReview ↗

Author response to Reviewer nhYX

We thank the reviewer for the valuable feedback. We address the weaknesses and questions below and we will incorporate them in the paper.

>>W1/Q1.

We thank the reviewer for raising this point. We agree that PixCalibrate is closely related to CalibrateMix and this is why we carefully compared them in the paper. PixCalibrate methodological contribution centers on the design of a new augmentation technique, which naturally resembles test-time corrupted images. Specifically, while CalibrateMix directly interpolates pixel values across two images, i.e., all contiguous pixels of an image are overlaid with all contiguous pixels from another image, according to the mixing ratio (see Fig. 2), PixCalibrate samples (uniformly at random) individual pixels from an image and places them into an anchor image, producing spatially distributed perturbations while retaining much of the anchor's structure (see also Fig. 2), especially because we preserve 60% of the anchor's pixels intact and replace only 40%. These 40% are not contiguous pixels to create an entire patch of pixels being replaced. Table 3 further supports this distinction, showing higher structural similarity to corruption perturbations for PixCalibrate (SSIM: 0.36) than Mixup (0.20) and CutMix (0.27).

In addition, we incorporate difficulty-aware pairing, which is inspired from CalibrateMix. However, while CalibrateMix combines easy labeled samples with hard pseudo-labeled samples, and hard labeled samples with easy pseudo-labeled samples, PixCalibrate uses a fixed direction to the construction: the easy sample provides the anchor, while the hard sample provides the perturbation. This design aims to preserve the features of the anchor while introducing challenging variations from the hard sample. As demonstrated in Table 5, reversing this direction by using hard samples as anchors can produce more ambiguous augmented samples, reducing label reliability. At our default replacement ratio (ρ=0.4\rho=0.4), 60% of spatial locations retain the easy image’s original pixel values, while the remaining locations receive pixels from the hard image. Thus, difficulty determines both which samples are paired and which sample acts as anchor and which sample acts as the perturbation source.

Importantly, these methodological differences contribute to improvements beyond those provided by confidence regularization. Although the confidence penalty complements our pixel-sampling strategy by encouraging higher predictive entropy, removing it still yields substantial improvements over CalibrateMix (Tables 1 and 5):

MethodCifar-10c accCifar-10c eceCifar-100c accCifar-100c ece
CalibrateMix75.5411.7751.6018.01
PixCalibrate without confidence penalty83.238.4363.0110.15
PixCalibrate84.707.6465.038.88

Furthermore, Appendix A.6.2 applies the same confidence penalty to Mixup and CutMix; these variants underperform PixCalibrate without the confidence penalty on both accuracy and ECE. This supports treating confidence regularization as a complementary component rather than the primary explanation for the gains.

We will revise the presentation to distinguish the inherited ingredients from the proposed construction. Specifically, our contribution is the integration of difficulty-dependent anchor/perturbation roles with spatially distributed pixel-sampling for jointly improving robustness and calibration under corruption. The pixel-sampling and replacement operator is central to this contribution, and the role-selection ablation provides evidence for the additional design choice beyond mixing granularity.

AuthorsOpenReview ↗

Author response to Reviewer nhYX

>>W2/Q2. Isolating Pixel-Level Mixing and CalibrateMix Adaptation to the Fully Supervised Setting

We thank the reviewer for this question and provide additional clarification regarding the comparison with global mixing and the fully supervised adaptation of CalibrateMix.

Comparison with Global Mixing. To further assess the contribution of pixel-level mixing, we evaluated CalibrateMix with the same AUM setting, under matched backbone and training settings as PixCalibrate, and with confidence penalty. We report the results in the table below:

MethodCifar10c eceCifar10c acc
CalibrateMix + Confidence Penalty9.8975.35
PixCalibrate7.6484.70

As shown from the table above, PixCalibrate achieves 84.70% accuracy and 7.64 ECE on CIFAR-10-C, compared with 75.35% accuracy and 9.89 ECE for CalibrateMix + Confidence Penalty, improving accuracy by 9.35 percentage points while reducing ECE by 2.25 points. Furthermore, Table 12 in Appendix shows that PixCalibrate without CP already achieves 83.23% accuracy and 8.43 ECE, outperforming Mixup with CP (75.14% / 10.47).

Fully Supervised Adaptation of CalibrateMix. We preserve CalibrateMix's original AUM-based difficulty estimation, top-k feature-dissimilarity selection, bidirectional pairing strategy, and global Mixup operation. In the fully supervised setting, since all training samples are labeled, the original easy-labeled–hard-unlabeled and easy-unlabeled–hard-labeled pairings are adapted to easy-labeled–hard-labeled pairings in both directions, using ground-truth labels instead of pseudo-labels. The SSL-specific unsupervised objectives are omitted (since no unlabeled data is available), while the remaining mixing loss is retained. All methods follow the same fully supervised training protocol. We will clarify these adaptation details in the revised manuscript.

W3/Q3. Similarity with Real-World Corruptions and Soft-Label Justification.

As illustrated in Fig. 2 in the paper, unlike Mixup, which globally mixes two images, and ResizeMix, which replaces contiguous regions, PixCalibrate introduces scattered pixel level changes while retaining the recognizable structure of the original image, resembling the fine-grained distortions observed in Gaussian noise and pixelation. In Section 6.1 and Table 3, we compare how closely our pixel-sampling augmentation resembles common corruptions. Specifically, Table 3 provides evidence that further supports this observation, showing that PixCalibrate achieves higher similarity score SSIM (0.36) than Mixup (0.20) and CutMix (0.27). Although PixCalibrate does not exactly reproduce real-world corruptions, these results suggest greater structural similarity to the evaluated corruption perturbations.

Regarding the labels, we use the pixel-replacement ratio as a mixing coefficient, following the area-proportional labeling principle of CutMix.

We thank the reviewer for suggesting the exploration of alternative labeling strategies. To investigate this, we conducted an additional experiment in which the original label of the anchor sample is preserved during mixing, rather than using pixel-proportion-based soft labels. The results are reported in the table below:

MethodCifar10c eceCifar10c acc
Mixup11.1575.41
Original anchor label7.8183.35
PixCalibrate7.6484.70

As shown in the table, preserving the anchor label achieves better calibration and accuracy than standard Mixup, suggesting that maintaining the anchor's semantic identity is beneficial. However, our proposed pixel-proportion-based soft labeling strategy further improves performance, reducing ECE from 7.81% to 7.64% and increasing accuracy from 83.35% to 84.70%. These results support the effectiveness of our labeling strategy in balancing semantic preservation with the contribution of mixed image content.

Dear Reviewer nhYX,

Thank you again for taking the time to review our paper and for your constructive feedback. We hope our responses and additional experiments adequately address your concerns. Please let us know if you have any remaining questions or would like further clarification.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.