- 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
- 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.
- 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.
- 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
- 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.
- 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.
- 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.
- 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
- What is the key innovative distinction between this work and Wang et al. (2024)?
- Could the authors verify and explain the differences in the transfer-attack results reported in Tables 2 and 7?
- How does the method perform when the disparity estimates used to establish correspondences contain substantial errors?
3 comments
Thank you for your response. Your rebuttal has largely addressed my concerns, and I will maintain my current positive rating of 3.
Thanks!
Thanks for your quick reply. We are happy that our rebuttal has largely addressed your concerns!
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, 20 iterations.
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) (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, 600 queries).