acceptodds
Under review as a conference paper at ICLR 2027

EpiMirror: Epipolar Correspondence Mirrored Perturbations Against Stereo Matching

Abstract

To understand the robustness of stereo-based 3D perception, adversarial perturbations against stereo matching have been increasingly studied. However, existing attacks do not exploit any specific vulnerability in this task, but simply follow conventional attacks (originally developed against image classification) that target final prediction errors or internal feature discrepancies. In this paper, we address this limitation by instead targeting epipolar structure, the task-specific geometric constraint on stereo correspondence. Specifically, we propose **EpiMirror**, a new attack that constructs **Mirror**ed perturbations along the **Epi**polar line, i.e., perturbations with pixel-wise opposite directions (but identical amplitude) for left vs. right views. In this way, EpiMirror explicitly disrupts cross-view matching, subsequently promoting disparities along the same epipolar line, and ultimately leading to erroneous depth estimation. Experimental results across six stereo architectures and four datasets show that EpiMirror improves the D1-error of existing white-box, black-box transfer-based, and black-box query-based attacks by (relative) 16.9%, 20.2%, and 79.4% on average. We further provide a theoretical analysis attributing the success of mirrored perturbations to amplification of correspondence mismatch and collapse of matching margin.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
3D perception, stereo matching, adversarial attacks, image perturbations

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

This paper proposes a geometry-aware adversarial attack on binocular stereo matching. By exploiting epipolar correspondences, it applies opposite-sign perturbations at corresponding locations in the left and right views to disrupt cross-view matching consistency. The method is compatible with white-box, black-box transfer, and query-based attacks, and achieves strong attack performance across multiple models and datasets.

Critical strengths
  1. The motivation is clear and task-specific. The method uses epipolar geometry to structure adversarial perturbations, making it more directly tailored to stereo matching than generic attacks.
  2. The method is simple and general. Opposite-sign coupling at corresponding locations does not require modifying the original attack objective and can be combined with different optimizers.
  3. The experimental evaluation is comprehensive and reasonably thorough, covering six stereo-matching architectures, four datasets, and three attack settings. Targeted ablations on correspondence, sign coupling, and the underlying mechanism support the effectiveness of the core design.
Critical weaknesses
  1. Wang et al. (2024) also attack stereo networks by increasing feature discrepancies between the left and right views. The substantive distinction between the proposed opposite-sign coupling and this prior approach needs further clarification, and a direct comparison is missing.
  2. The analysis relies on local first-order approximations and specific alignment conditions. The matching-margin result is primarily an algebraic condition and does not sufficiently explain how disparity errors arise in practice.
  3. The patch-based transfer-attack results in Table 2 of the main paper and Table 7 of the appendix cannot be directly reconciled. These results should be checked and explained.
  4. The method relies on a fixed disparity field to establish correspondences, but its robustness to inaccurate correspondences is not sufficiently discussed.

Wang et al. (2024): Left-right Discrepancy for Adversarial Attack on Stereo Networks

Questions for authors
  1. What is the key innovative distinction between this work and Wang et al. (2024)?
  2. Could the authors verify and explain the differences in the transfer-attack results reported in Tables 2 and 7?
  3. How does the method perform when the disparity estimates used to establish correspondences contain substantial errors?
3 comments

W1&Q1. Comparison to Wang et al. (2024).

New results in Table R4 confirms that EpiMirror outperforms Wang et al. (2024). In the Abstract, we state that “existing attacks do not exploit any specific vulnerability in this task, but simply follow conventional attacks (originally developed against image classification) that target final prediction errors or internal feature discrepancies.” Wang et al. (2024) belong to them and specifically rely on image-level internal feature discrepancies. In contrast, we target the task-specific geometric constraint on stereo correspondence and rely on pixel-level vulnerability with our mirrored perturbations. Note that Wang et al. (2024) was just posted on arXiv in 2024-Jan., without further publication or updates.

Table R4. White-box comparisons on KITTI 2015. RAFT-Stereo, 40 pairs, ϵ=4/255\epsilon=4/255, 20 iterations.

MethodD1 (%) ↑EPE (px) ↑
Clean3.590.67
Wang et al. (2024)30.523.04
EpiMirror43.034.70

W2. From Matching-Cost Disruption to Disparity Errors.

We clarify that Proposition 2 characterizes the following preference shift within an effective matching-cost formulation, rather than providing a complete guarantee for nonlinear stereo networks. Our first-order analysis characterizes local mismatch amplification under the stated alignment condition, rather than guaranteeing global attack behavior. In practice, EpiMirror introduces opposite-sign perturbations at estimated correspondences, disrupting cross-view agreement and potentially increasing the matching cost of the original correspondence relative to competing disparities. When this relative cost change exceeds the original matching margin, a competing disparity becomes preferable. Such changes may propagate through cost aggregation and iterative refinement, ultimately leading to erroneous disparity predictions.

W3&Q2. Copy-paste errors in Tables 2 and 7.

We sincerely apologize for them and correct them in the revised PDF.

W4&Q3. Impact of correspondence accuracy on attack performance of EpiMirror.

The clean results in Table 1 of the PDF has already confirmed that our tested six models are accurate enough, with small EPEs in the range of [0.44, 2.13]. New results in Table R5 further show that using the most accurate, ground-truth disparity only yields small attack improvement (D1: 52.59% → 54.08%, EPE: 44.30 → 45.21). Reversely, if we intentionally introduce more correposnce errors using dλ=dGT+λ(dpred−dGT)d_{\lambda}=d_{\mathrm{GT}}+\lambda(d_{\mathrm{pred}}-d_{\mathrm{GT}}) (not practically meaningful but only for testing), the attack results get worse as expected.

Table R5. Impact of correspondence accuracy on attack performance of EpiMirror. Results are evaluated on Middlebury using RAFT-Stereo (ϵ=4/255\epsilon=4/255, 600 queries).

DisparityCorrespondence Error (px) ↓Adversarial D1 (%) ↑Adversarial EPE (px) ↑
Ground Truth (λ=0\lambda=0)0.0054.0845.21
Predicted (ours) (λ=1\lambda=1)1.1152.5944.30
Degraded: λ=2\lambda=22.2251.1649.13
Degraded: λ=4\lambda=44.4442.7335.83
Degraded: λ=6\lambda=66.6643.4938.23
Degraded: λ=8\lambda=88.8840.9832.55
Degraded: λ=10\lambda=1011.1038.7233.16

Thank you for your response. Your rebuttal has largely addressed my concerns, and I will maintain my current positive rating of 3.

Thanks for your quick reply. We are happy that our rebuttal has largely addressed your concerns!

Reviewer_FQX6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes EpiMirror, a geometry-aware adversarial attack for stereo matching. It uses a fixed disparity field to apply opposite-sign perturbations at corresponding locations in the left and right views, supporting white-box, transfer-based black-box, and query-based black-box attacks. The authors explain its mechanism through local theoretical analysis and validate its effectiveness through comparisons and ablations across multiple models and datasets.

Critical strengths
  1. The method is well motivated and uses a simple mechanism: applying opposite-sign perturbations at stereo correspondences.
  2. The experiments cover multiple stereo matching models, datasets, and attack settings. The Same, Pixel, Shuffled, and Independent variants help examine the roles of different design choices.
Critical weaknesses
  1. EpiMirror relies on placing opposite-sign perturbations at corresponding locations in the left and right images. These locations are derived from predicted disparity and may therefore be inaccurate. The paper mainly compares correspondence-based pairing with shuffled pairing and other variants, but does not systematically examine how attack performance changes as correspondence accuracy degrades.
  2. Several important implementation details are insufficiently described. Equation (1) introduces the attack score SS and scoring region MM, but neither is explicitly defined. The rendering operator also lacks details on interpolation, occlusion handling, and how multiple left-view pixels mapping to the same right-view location are handled.
  3. Some differences in the reported results are not explained. For example, the EpiMirror query-based attack against RAFT-Stereo on KITTI achieves D1/EPE = 38.57/27.36 in Table 3, compared with 31.09/14.59 in Table 4.Table 7 also contains identical numerical rows under different method labels: the EpiMirror row with IGEV-Stereo as the source model in the full-image attack group is identical to the DepthVanish row with the same source model in the patch-based attack group.
Questions for authors
  1. Provide the exact definitions of the attack score SS and scoring region MM in Equation (1). How is the scoring region determined, and are the same scoring rules used across different attack methods and settings?
  2. How is the rendering operator WdgeoW_{d_{\mathrm{geo}}} implemented? Specify the interpolation scheme for non-integer disparities and the rules for handling occlusions, out-of-bounds mappings, and multiple left-view pixels mapping to the same right-view location.
  3. After cross-view mapping, interpolation, and pixel clipping, do perturbations at corresponding locations still retain opposite signs and equal magnitudes?
  4. Verify and explain the repeated experimental results reported under different methods in Table 7.
5 comments

W1. Impact of correspondence accuracy on attack performance of EpiMirror.

The clean results in Table 1 of the PDF has already confirmed that our tested six models are accurate enough, with small EPEs in the range of [0.44, 2.13]. New results in Table R2 further show that using the most accurate, ground-truth disparity only yields small attack improvement (D1: 52.59% →\rightarrow 54.08%, EPE: 44.30 →\rightarrow 45.21). Reversely, if we intentionally introduce more correposnce errors using dλ=dGT+λ(dpred−dGT)d_{\lambda}=d_{\mathrm{GT}}+\lambda(d_{\mathrm{pred}}-d_{\mathrm{GT}}) (not practically meaningful but only for testing), the attack results get worse as expected.

Table R2. Impact of correspondence accuracy on attack performance of EpiMirror. Results are evaluated on Middlebury using RAFT-Stereo (ϵ=4/255\epsilon=4/255, 600 queries).

DisparityCorrespondence Error (px) ↓Adversarial D1 (%) ↑Adversarial EPE (px) ↑
Ground Truth (λ=0\lambda=0)0.0054.0845.21
Predicted (ours) (λ=1\lambda=1)1.1152.5944.30
Degraded: λ=2\lambda=22.2251.1649.13
Degraded: λ=4\lambda=44.4442.7335.83
Degraded: λ=6\lambda=66.6643.4938.23
Degraded: λ=8\lambda=88.8840.9832.55
Degraded: λ=10\lambda=1011.1038.7233.16

W2&Q1. Details of Attack Score and Scoring Region.

We clarify the definition of attack score, scoring region, and their consistency across settings in the revised PDF.

Attack Score. For output-only black-box attacks, we use the mean absolute disparity deviation:

S(dadv,d0;M)=1∣M∣∑p∈M∣dadv(p)−d0(p)∣,S(d_{\mathrm{adv}},d_0;M)=\frac{1}{\lvert M\rvert}\sum_{p\in M}\lvert d_{\mathrm{adv}}(p)-d_0(p)\rvert,

where d0d_0 and dadvd_{\mathrm{adv}} denote the clean and adversarial predictions, respectively. This score requires no ground truth and guides perturbation optimization.

Scoring Region. We define a fixed scoring mask:

M=Ω∩Vclean∩Vgeo∩Vproj,M=\Omega\cap V_{\mathrm{clean}}\cap V_{\mathrm{geo}}\cap V_{\mathrm{proj}},

where Ω\Omega denotes the full left image, while VcleanV_{\mathrm{clean}} and VgeoV_{\mathrm{geo}} exclude invalid clean predictions and geometry disparities, respectively. VprojV_{\mathrm{proj}} retains pixels with at least one valid in-bounds projection, using horizontal linear splatting for non-integer disparities and discarding out-of-bounds contributions. The mask remains fixed throughout optimization, independent of adversarial predictions or GT correctness.

Consistency Across Settings. All matched output-only black-box attacks use the same attack score, fixed scoring region, and query budget, while differing in their optimization strategies. When ground truth is used for scoring, the region is additionally restricted to clean-correct pixels. White-box and transfer-based methods retain their respective optimization objectives. All methods are evaluated using the same D1/EPE protocol.

W2&Q2. Details of Rendering Operator.

We implement Wdgeo\mathcal{W}_{d_{\mathrm{geo}}} using normalized horizontal forward splatting with fixed disparities. For non-integer disparities, we compute xR=x−dgeo(x,y)x_R=x-d_{\mathrm{geo}}(x,y) and a=xR−⌊xR⌋a=x_R-\lfloor x_R\rfloor, distributing perturbations to (⌊xR⌋,y)(\lfloor x_R\rfloor,y) and (⌊xR⌋+1,y)(\lfloor x_R\rfloor+1,y) with weights 1−a1-a and aa, respectively. Invalid disparities and out-of-bounds contributions are discarded individually without clamping. To handle occlusions and many-to-one mappings, a disparity-based z-buffer retains the largest disparity (nearest estimated surface), discarding smaller-disparity contributions. Contributions within 10−610^{-6} px of the maximum disparity are combined by normalized weighted averaging. This approximates visibility without explicit occlusion detection.

Q3. Retaining opposite signs and equal magnitudes.

Table R3 shows that 87.78% of pixels retain both, and only 1.29% violate both. Here, we selected 5 previously saved stereo pairs from the Middlebury training set (Jadeplant, MotorcycleE, Playroom, Playtable, and Recycle). We extract the actual perturbations directly from the saved clean and adversarial images, with clipping and quantization effects. Using the original rendering operator, we check

∣δRactual(q)+(WdgeoδLactual)(q)∣≤10−7\left|\delta_R^{\mathrm{actual}}(q) + (\mathcal{W}_{d_{\mathrm{geo}}}\delta_L^{\mathrm{actual}})(q)\right|\leq 10^{-7}

for each pixel. A pixel violates the coupling condition if any RGB channel exceeds this tolerance. We evaluate only right-view pixels with projection support, pooling 1,499,226 pixels across the five final adversarial pairs.

Table R3. Ratios of pixels retaining opposite signs and equal magnitudes after rendering. Evaluated using RAFT-Stereo under the output-only black-box setting, with ϵ=4/255\epsilon=4/255, 600 queries per pair.

Opposite SignEqual MagnitudePixelsRatio (%)
××19,3251.29
×✓133,3228.89
✓×30,5312.04
✓✓1,316,04887.78

W3&Q4. Errors in Tables 3, 4 and 7.

We sincerely apologize for them and address them in the revised PDF. Specifically, we correct the copy-paste errors in Table 7 and clarify that the ablation study in Table 4 was conducted on a subset of 15 KITTI stereo pairs, different from Table 3.

Thank you for the detailed rebuttal and additional experiments. The authors have addressed most of my original concerns, particularly regarding correspondence accuracy, the rendering operator, and the experimental inconsistencies.

However, two issues remain:

  1. Preservation of the mirrored constraint. The reported results show that only 87.78% of pixels simultaneously retain opposite signs and equal magnitudes after rendering and input processing. It would be helpful to clarify whether this proportion is consistent across individual stereo pairs and whether the remaining violations affect attack performance.

  2. Consistency of optimization and evaluation. Although the attack score and scoring region are now clarified, different attack settings retain different optimization objectives. A clearer explanation of the scoring regions and optimization protocols would improve the fairness and reproducibility of the comparisons.

Overall, the rebuttal has addressed several important concerns and improved the manuscript. However, the above points still deserve clarification.

Reviewer_gL76OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes an attack named EpiMirror against stereo matching. The key novelty is that it constructs mirrored perturbations along the epipolar line. In this way, the method explicitly disrupts cross-view matching. The method is validated on six stereo architectures and four datasets, and the proposed method improves the white-box, black-box transfer-based, and black-box query-based attacks.

Critical strengths
  1. The idea is clear and reasonable. Utilizing the epipolar geometry for a stereo matching attack is genuinely task-specific, unlike attacks borrowed from classification.

  2. The method is simple and composable with near-zero overhead. Compared with existing methods, no new loss, no extra parameters, no access to intermediate features, and no change to the optimizer.

  3. The query-based black-box gain is convincing.

Critical weaknesses
  1. Critical experiemntal issues. Tables 2, 6, and 7 contain apparent copy-paste errors, with Table 7's duplicated EpiMirror rows reusing Stereopagnosia's D1 and EpiMirror's EPE, and Table 6 reporting absurd EPE values above 400 px: EPE shows values such as 416.32 / 308.37 / 376.93 px. A perturbation of ε = 4/255 cannot possibly cause a parallax error exceeding 400 pixels; it is almost certainly due to a bug in normalization or unit conversion. The data in this table is currently unreliable. Moreover, in Table 2, the EPE of the patch-based EpiMirror row is exactly the same as that of the full-image row (AANet 4.62 / RAFT 7.20 / IGEV 6.92 / STTR 14.44), but D1 is different (RAFT patch 27.83 < full 36.05, but the EPE is the same), presenting an internal contradiction.

  2. Limited novelty. Cheng et al. (2022a) have already constructed photometrically consistent perturbations at the corresponding positions, and Wang et al. (2024) have already identified discrepancies in the geometric alignment features. Both of these have accomplished the step of "matching perturbations at corresponding positions".

  3. The motivation of the paper states "Understanding the robustness of 3D perception based on stereo", but no defenses were tested.

  4. There is no verification from the physical world. The references Zhao 2026, Xing 2025, and Liu 2025 all conducted physical/ printable attacks. This paper stops at digital perturbations. The actual threat in the autonomous driving scenario has not been confirmed.

Questions for authors
  1. Critical experiemntal issues. Tables 2, 6, and 7 contain apparent copy-paste errors, with Table 7's duplicated EpiMirror rows reusing Stereopagnosia's D1 and EpiMirror's EPE, and Table 6 reporting absurd EPE values above 400 px: EPE shows values such as 416.32 / 308.37 / 376.93 px. A perturbation of ε = 4/255 cannot possibly cause a parallax error exceeding 400 pixels; it is almost certainly due to a bug in normalization or unit conversion. The data in this table is currently unreliable. Moreover, in Table 2, the EPE of the patch-based EpiMirror row is exactly the same as that of the full-image row (AANet 4.62 / RAFT 7.20 / IGEV 6.92 / STTR 14.44), but D1 is different (RAFT patch 27.83 < full 36.05, but the EPE is the same), presenting an internal contradiction.

  2. Limited novelty. Cheng et al. (2022a) have already constructed photometrically consistent perturbations at the corresponding positions, and Wang et al. (2024) have already identified discrepancies in the geometric alignment features. Both of these have accomplished the step of "matching perturbations at corresponding positions".

  3. The motivation of the paper states "Understanding the robustness of 3D perception based on stereo", but no defenses were tested.

  4. There is no verification from the physical world. The references Zhao 2026, Xing 2025, and Liu 2025 all conducted physical/ printable attacks. This paper stops at digital perturbations. The actual threat in the autonomous driving scenario has not been confirmed.

6 comments

W1&Q1. Table results.

(1) Copy-paste errors in Tables 2 and 7. We sincerely apologize for them and correct them in the revised PDF.

(2) Misunderstanding about Table 6. We clarify that ϵ=4/255\epsilon=4/255 constrains only input pixel perturbations, not output disparity prediction errors, so EPE values above 400 px are possible. The general philosophy of adversarial examples is that small input perturbations induce large output differences.

W2&Q2. Novelty of “matching perturbations at corresponding positions”.

We clarify that Wang et al. (2024) and Cheng et al. (2022) have NOT done so. Specifically, Wang et al. (2024) consider only image-level feature discrepancies and do not consider pixel-level positions. Cheng et al. (2022) simply apply the same perturbations to the left and right views, which is completely irrelevant to our uncovered “mirrored vulnerability”. Note that Wang et al. (2024) was posted on arXiv in 2024-Jan, and has not been published or updated afterwards.

Both studies fall into what we have stated in the Abstract: “existing attacks do not exploit any specific vulnerability in this task, but simply follow conventional attacks (originally developed against image classification) that target final prediction errors or internal feature discrepancies.” In contrast, we target the task-specific geometric constraint on stereo correspondence and rely on pixel-level vulnerability with our mirrored perturbations.

We further give a theoretical justification for the essential difference between Cheng et al. and our EpiMirror. Let Δ=IL(p)−IR(q∗)\Delta=I_L(p)-I_R(q^{\ast}) and Δ′=IL(p)−IR(q′)\Delta'=I_L(p)-I_R(q') denote the intensity differences for correct and incorrect matches, respectively. Let u,u′u,u' denote the corresponding left-view perturbations, with ∣u∣,∣u′∣≤ϵ\lvert u\rvert,\lvert u'\rvert\leq\epsilon. Under idealized pixel-wise absolute-difference matching with exact correspondences, we have

C_{\mathrm{Cheng}}(d^{\ast}) &= \lvert\Delta\rvert, & C_{\mathrm{Cheng}}(d') &= \lvert\Delta'+u-u'\rvert, \\\\ C_{\mathrm{Epi}}(d^{\ast}) &= \lvert\Delta+2u\rvert, & C_{\mathrm{Epi}}(d') &= \lvert\Delta'+u+u'\rvert. \end{aligned}$$ Our Proposition 2 in the PDF identifies two pathways for reducing the matching-cost margin and making an incorrect disparity preferable: $$M_{\mathrm{adv}} = M_0 + \underbrace{\Delta C(d')}_{\substack{\textbf{Pathway 1:}\\ \text{Decrease incorrect cost} \\\\ \text{Cheng \\& EpiMirror}}} - \underbrace{\Delta C(d^{\ast})}_{\substack{\textbf{Pathway 2:}\\ \text{Increase correct cost} \\\\ \text{EpiMirror only}}}.$$ Cheng et al.'s same-sign coupling preserves the correct-match cost ($\Delta C(d^{\ast})=0$), restricting its matching-margin reduction to Pathway 1. In contrast, EpiMirror's opposite-sign coupling can exploit both pathways, leading to stronger attacks, consistent with the mismatch amplification characterized by Proposition 1 under its alignment condition. Our ablation results in Table 4 support the superiority of our opposite-sign perturbations. ### **W3&Q3. Defense results.** New results in Table R1 with four low-cost, pre-processing-based defenses show that our EpiMirror remains the best. In particular, we consider an *adaptive defense* against our mirrored perturbations: recovering the mirrored perturbations as $\hat{\delta}(p)=(I_L^{adv}(p)-I_R^{adv}(q))/2$, where $q=(x-\hat{d}(p),y)$ is determined by predicted disparity, assuming approximate photometric consistency. However, even under such an adaptive attack, EpiMirror remains effective. This may be because the approximated photometric consistency does not hold well. We leave detailed exploration of more targeted defenses for future work. **Table R1. White-box Attack results (Stereopagnosia/Stereoscopic/EpiMirror) under four low-cost, pre-processing-based defenses and one EpiMirror-adapted defense.** Evaluated on KITTI 2015 using RAFT-Stereo ($\epsilon=4/255$). | Defense | Clean D1 (%) ↑ | Clean EPE (px) ↑ | Adversarial D1 (%) ↑ | Adversarial EPE (px) ↑ | |---|---:|---:|---:|---:| | None | 3.59 | 0.67 | 40.24 / 38.15 / **43.03** | 3.29 / 3.96 / **4.70** | | Gaussian ($\sigma=1.0$) | 3.79 | 0.69 | 33.75 / 27.35 / **39.78** | 2.49 / 2.14 / **2.93** | | JPEG ($Q=70$) | 4.37 | 0.77 | 30.25 / 24.46 / **41.85** | 2.33 / 1.97 / **3.04** | | Color jitter ($B/C/S=0.1,\\ H=0.02$) | 3.73 | 0.68 | 31.38 / 25.26 / **42.81** | 2.37 / 2.00 / **3.08** | | Adaptive | 3.61 | 0.70 | 33.31 / 25.82 / **42.71** | 2.45 / 2.05 / **3.09** |

W4&Q4. Physical-world experiments.

We acknowledge that physical-world experiments are commonly done for patch-based attacks, which apply unconstrained perturbations. In contrast, imperceptible perturbation-based attacks, including Stereopagnosia and Stereoscopic, and ours, focus on the digital world. In particular, we have already shown that the patch-based variant of our EpiMirror outperforms the patch-based DepthVanish by 6.09% in D1 and 26.04% in EPE on average (see Table 1 in the PDF), suggesting its good potential in the physical world.

Thanks for the response. In my opinion, imperceptible perturbation-based attacks can hardly achieve physical effectiveness without specific designs, which cannot be addressed in the current version. Without a physical implementation, I feel the attack is not well motivated other wise the work can demonstrate the digital attacks could help enhance the robustness of the depth estimation.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions to the Manuscript

Summary of Revisions to the Manuscript. We have revised the manuscript to address the reviewers' concerns. All revisions are marked in blue in the revised PDF.

  1. Novelty clarification. Clarified the key differences between EpiMirror and the methods of Cheng et al. (2022a) and Wang et al. (2024).
  2. Implementation details. Specified the definitions of the scoring function (SS), scoring mask (MM), and rendering operator.
  3. Ablation clarification. Clarified the dataset and experimental settings used in Table 4.
  4. Defense evaluation. Added evaluations of preprocessing defenses and an EpiMirror-adapted defense.
  5. Correspondence robustness. Added experiments assessing the sensitivity of EpiMirror to correspondence errors.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.