acceptodds
Under review as a conference paper at ICLR 2027

MambaX‑AD: Region‑Aware Contrastive Learning with Mamba-CNN Cross‑Modulation for Multimodal Anomaly Detection

Abstract

In industrial manufacturing, multimodal anomaly detection combining 2D images and 3D point clouds provides complementary appearance and geometric cues for defect inspection. However, existing unsupervised cross-modal mapping methods typically rely on passive fitting of normal correspondences without explicit negative constraints, which can leave an overly broad normal feature boundary and weak discriminative sensitivity. Moreover, static cross-modal mappings have limited capacity to reconcile the asymmetric local details and global context of the two modalities, resulting in a cross-modal alignment bottleneck. To address these two limitations, we propose MambaX-AD. First, we introduce region-aware contrastive learning (RA-CL), a region-aware, mask-conditioned contrastive strategy that uses the spatial mask of feature-level pseudo-anomalies to impose stronger repulsion on corrupted regions while applying a mild balancing penalty to uncorrupted nominal regions. Second, we introduce reciprocal Mamba-CNN cross-modulation mechanism, in which the local CNN and global Mamba branches generate modulation signals for each other and direction- and depth-aware embeddings condition these interactions. Experiments on MVTec 3D-AD achieve the highest mean image-level I-AUROC among the compared methods while maintaining competitive localization performance, and experiments on Eyecandies achieve the highest results on I-AUROC, P-AUROC, and AUPRO@1% among the compared methods.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
multimodal anomaly detection, cross-modal mapping

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_qdyJOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MambaX-AD for unsupervised multimodal anomaly detection, extending cross-modal feature mapping with a region-aware contrastive learning based on feature-level pseudo-anomalies and a reciprocal Mamba-CNN cross-modulation mechanism for 2D–3D feature alignment. The method is evaluated on MVTec 3D-AD and Eyecandies, where it reports good image-level anomaly detection performance and competitive localization results.

Critical strengths

The experimental evaluation is relatively comprehensive. The paper evaluates the proposed method on both MVTec 3D-AD and Eyecandies, and includes component-wise ablations, multi-seed experiments, hyperparameter sensitivity analysis, and qualitative comparisons, providing reasonably broad empirical validation of the method.

Critical weaknesses
  1. Pseudo anomalies are constructed only on the target features while the cross-modal mapper still receives clean source features, making RA-CL closer to output-space margin regularization. It remains unclear why such target-side pseudo-negative repulsion should improve sensitivity to real anomalous inputs at test time.
  2. A nonzero negative weight αbf\alpha_{\mathrm{bf}} is still assigned to uncorrupted regions and is justified as mitigating receptive-field leakage, although the pseudo anomalies are generated after feature extraction. If Ei−=Ei+E_i^- = E_i^+ at these locations, the same normal feature is simultaneously subject to the positive alignment term and the negative margin penalty, resulting in conflicting optimization objectives.
  3. The claimed benefits of region-aware learning and local-global cross-modulation are not well supported by the ablation results, since the improvements in anomaly localization remain very limited and can even degrade when RA-CL replaces standard contrastive learning.
Questions for authors
  1. In RA-CL, pseudo anomalies are constructed only on the target features, while the cross-modal mapper still receives clean source features. Why can such a target-side pseudo-negative constraint improve sensitivity to real anomalous inputs at test time? In addition, when Ei−=Ei+E_i^- = E_i^+ at uncorrupted locations, why is the negative margin penalty weighted by αbf\alpha_{\mathrm{bf}} still applied?
  2. Table C.3 reports a higher inference FPS for MambaX-AD than for CFM despite introducing additional Mamba-CNN branches, reciprocal modulation, and feature fusion. Were CFM and MambaX-AD evaluated under exactly the same preprocessing, backbone, feature resolution, numerical precision, and implementation environment, and what specifically leads to the higher throughput?
5 comments
AuthorsOpenReview ↗

Why target-side pseudo-anomalies improve real anomaly detection.

We sincerely thank the reviewer for the constructive comments and insightful questions regarding the target-side pseudo-anomaly construction, background negative weighting, and the effectiveness of our proposed mechanisms. We address each concern below.

We would like to clarify that target-side pseudo-anomaly construction is directly motivated by our cross-modal prediction-based anomaly scoring mechanism.

For the 2D-to-3D mapping, the model takes a clean 2D feature as input and predicts its corresponding 3D feature. We construct the positive target using the original clean 3D feature and the negative target by perturbing this feature in the target space. The same procedure is applied symmetrically to the 3D-to-2D mapping.

The key reason for perturbing the target rather than the source is that our inference anomaly score is determined by the discrepancy between the predicted target feature and the observed target feature. By introducing pseudo-anomalous negatives directly in the target space, RA-CL explicitly encourages the cross-modal prediction to distinguish clean normal targets from perturbed target representations. Meanwhile, keeping the source feature clean preserves the normal cross-modal conditioning relationship and avoids introducing additional uncertainty from source-side corruption.

Importantly, these pseudo-anomalies are not intended to reproduce specific real-world defects. Instead, they provide controlled feature-space perturbations around normal target representations, encouraging the model to become more sensitive to deviations in the same feature space used for inference.

AuthorsOpenReview ↗

Why retain the negative penalty at uncorrupted locations.

We acknowledge the reviewer's observation that the positive and negative targets are identical at uncorrupted locations. Consequently, the two terms introduce competing optimization effects: the positive term encourages alignment with the clean normal target, while the negative margin term prevents excessively tight point-wise alignment within the prescribed margin.

We intentionally retain this term as a background soft-margin regularization, rather than treating uncorrupted locations as actual anomalies. Its purpose is to balance normal feature alignment and discriminative separation.

To verify its practical effectiveness, we conducted a sensitivity experiment under the complete MambaX-AD architecture, fixing αlocal=3\alpha_{local}=3 and γ=2\gamma=2, while varying only αbf\alpha_{bf}. All configurations use the same random seed and identical remaining settings.

αbf\alpha_{bf}I-AUROCAUPRO@1%
00.9760.448
10.9820.456
30.9740.448

Removing the background negative term (αbf=0\alpha_{bf}=0) degrades both detection and localization performance, while increasing its weight to 3 also leads to worse results. The default value of 1 achieves the best performance among the tested configurations.

These results demonstrate that a moderate background negative weight is beneficial in practice, whereas removing it or applying excessively strong repulsion is detrimental. We will clarify this optimization behavior in the revised manuscript.

AuthorsOpenReview ↗

Effectiveness of the proposed RA-CL mechanism.

We thank the reviewer for examining the effectiveness of our two proposed mechanisms and raising concerns about their ablation results.

Regarding RA-CL, under the same multi-scale Mamba-CNN configuration, replacing Standard CL with RA-CL changes I-AUROC from 0.977 to 0.978, while AUPRO@1% decreases from 0.455 to 0.449. However, this comparison is conducted under an intermediate M-CNN architecture without reciprocal cross-modulation and should not be interpreted as the primary evidence for the independent effectiveness of RA-CL or the performance of our complete architecture.

To independently evaluate RA-CL, we conducted a controlled comparison without our second contribution. Standard CL achieves 0.972/0.455, whereas RA-CL achieves 0.974/0.455 (I-AUROC/AUPRO@1%). This demonstrates that our region-aware formulation provides additional image-level discriminative benefits while maintaining low-FPR localization performance. More importantly, introducing contrastive learning into normal-only cross-modal prediction constitutes the core of our first contribution, while region awareness further enhances this objective. This is also reflected in our staged ablation, where incorporating RA-CL after MS improves both metrics from 0.962/0.448 to 0.974/0.455. Thus, our first contribution encompasses not only region-aware weighting, but also the introduction of discriminative contrastive supervision, further enhanced through mask-conditioned pseudo-anomaly construction and region-aware weighting. Together, these results support the effectiveness of our first contribution as a whole.

Reviewer_AptNOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MambaX-AD for multimodal industrial anomaly detection using paired RGB images and 3D point clouds. The framework learns bidirectional mappings between spatially aligned, multi-scale features extracted by frozen backbones. Its two principal components are region-aware contrastive learning (RA-CL), which applies mask-conditioned negative penalties using feature-level pseudo-anomalies, and reciprocal Mamba–CNN cross-modulation, which couples local and global branches through gates conditioned on direction- and depth-aware embeddings. Anomaly scores are obtained from discrepancies between predicted and observed features. The method is evaluated on MVTec 3D-AD and Eyecandies, with additional component ablations, repeated-run results, and efficiency analyses.

Critical strengths
  1. The reported MVTec 3D-AD mean I-AUROC is 0.982, compared with 0.954 for CFM and 0.972 for MAFR in Table 1. On Eyecandies, the method achieves the highest reported I-AUROC, P-AUROC, and AUPRO@1% among the methods included in Table 2. These results provide meaningful empirical support for the complete system. The distinction between improved image-level detection and competitive, rather than uniformly superior, MVTec localization is also appropriate.
  2. The evaluation extends beyond a single main-results table. Appendix C.1 reports five-seed results across all ten MVTec categories, Appendix C.3 includes a controlled replacement of the Mamba branch with quadratic self-attention, and Appendix D reports an Eyecandies ablation using the same RA-CL coefficients. These experiments are valuable for assessing the repeatability and practical feasibility of the complete framework. In particular, the computational comparison directly examines the motivation for using a Mamba-based global branch rather than relying only on an asymptotic complexity argument.
Critical weaknesses
  1. According to Section 3.2 and Figure 3, the negative target is constructed by perturbing the clean target feature, and the positive and negative pairs share the same predicted feature. Unless an additional operation modifies locations outside the corruption mask, an uncorrupted location satisfies E⁻⁽ⁱ⁾ = E⁺⁽ⁱ⁾, implying dᵢ⁻ = dᵢ⁺ = d. Under Eqs. (2)–(5), its loss, up to a common normalization factor, is therefore ℓᵢ(d) = d² + α_bf max(0, m − d)². As a scalar function of d, this is minimized at d* = [α_bf / (1 + α_bf)] m. With the stated default α_bf = 1, this gives d* = m/2, rather than zero. Thus, the loss simultaneously attracts and repels the prediction relative to the same clean target at these locations. This calculation does not establish that the complete model fails or that every location reaches this residual. It does, however, raise a concrete concern about the claim that the term mitigates receptive-field leakage without disrupting normal alignment. The manuscript needs to identify any operation that invalidates the equality above, or explain why the induced nonzero normal residual is beneficial. The derivative analysis in Appendix C.2 concerns the negative term at corrupted locations and does not resolve this issue. This matters because preserving normal alignment while selectively repelling pseudo-anomalies is central to the proposed contribution.
  2. In Table 4, replacing Std-CL with RA-CL under the same multi-scale Mamba–CNN configuration changes I-AUROC from 0.977 to 0.978, while AUPRO@1% decreases from 0.455 to 0.449. This does not negate the image-level improvement, but it provides mixed evidence for the additional region-aware formulation. Crucially, the table does not compare Std-CL and RA-CL within the complete architecture containing cross-modulation and aware embeddings. Consequently, the final system’s stronger performance does not establish that RA-CL is preferable to standard contrastive learning in that system. The architectural evidence has a related limitation. Table 6 compares reciprocal modulation with direct addition and only one single-direction modulation variant. A capacity-comparable concatenation/projection baseline and the reverse single-direction variant would help distinguish the benefit of reciprocal information flow from that of additional learnable transformations. Furthermore, the five-seed results concern the complete model, not the key component comparisons, so they do not establish the repeatability of the individual gains. These gaps leave the contribution of the specific proposed mechanisms less firmly established than the effectiveness of the overall system.
Questions for authors
  1. At locations where M_pa = 0, are E⁻ and E⁺ identical in the actual implementation? If not, which operation modifies these locations after pseudo-anomaly generation? If they are identical, how should the nonzero preferred residual derived above be interpreted, and what evidence supports retaining this negative penalty rather than setting α_bf = 0? Please also clarify whether features are L2-normalized before computing the training distances, since training uses Euclidean distance while inference uses cosine distance.
  2. Do you have a comparison between Std-CL and RA-CL with the complete architecture held fixed, preferably using matched training settings and repeated seeds? For cross-modulation, do capacity-comparable concatenation/projection and reverse single-direction modulation controls support the same conclusion? These are requests for targeted additional evidence, rather than implementation clarification; their outcomes would materially affect my assessment of the contributions.
  3. Which data split and criterion are used for the periodic evaluations mentioned in Section 4.1, and how are the final checkpoints and hyperparameters selected? Are the main-table results from a designated run or aggregated across seeds? In addition, why is CFM omitted from the localization comparison in Table 3 despite being the closest mapping baseline and having a baseline AUPRO@1% of 0.455 in Table 4? Please clarify which baseline results are reproduced under the current protocol and which are taken from prior publications.
5 comments
AuthorsOpenReview ↗

RA-CL Optimization and Background Soft-Margin Regularization.

We sincerely thank the reviewer for the careful mathematical analysis and constructive suggestions. We particularly appreciate the positive assessment of our main benchmark results, multi-seed evaluation, cross-dataset analyses, and controlled efficiency experiments.

We thank the reviewer for the careful analysis of our RA-CL objective. We confirm that, at uncorrupted locations (Mpa(i)=0M_{pa}^{(i)}=0), the positive and negative targets are identical. Thus, as correctly derived by the reviewer, the per-location objective (omitting the common normalization factor) is

ℓi(d)=d2+αbf[m−d]+2\ell_i(d)=d^2+\alpha_{bf}[m-d]_+^2

with an idealized minimizer of

d∗=αbf1+αbfmd^*=\frac{\alpha_{bf}}{1+\alpha_{bf}}m

d∗=m2,when αbf=1d^*=\frac{m}{2},\quad \text{when }\alpha_{bf}=1

However, the fact is that this behavior was not sufficiently explained in the original manuscript. Importantly, this is an intentional design choice rather than an implementation artifact.

This behavior is a controlled background soft-margin regularization. The positive term encourages alignment with the clean normal target, while the negative margin term introduces a controlled repulsive constraint. Their balance prevents excessively tight point-wise fitting and establishes a finite residual distance regulated by (αbf\alpha_{bf}), without unbounded repulsion.

To verify its effectiveness, we conducted an additional sensitivity experiment under the complete architecture, fixing (αlocal=3\alpha_{local}=3) and (γ=2\gamma=2), and varying only (αbf\alpha_{bf}). All configurations use the same random seed and identical remaining settings.

αbf\alpha_{bf}I-AUROCAUPRO@1%
00.9760.448
10.9820.456
30.9740.448

Removing the background negative term (αbf=0\alpha_{bf}=0) reduces I-AUROC by 0.006 and AUPRO@1% by 0.008 relative to the default setting. Increasing its weight to 3 also degrades both metrics. Therefore, the moderate default weight of 1 achieves the best performance among the tested settings. These controlled results support the effectiveness of a nonzero but moderate background regularization, while also showing that stronger repulsion is not necessarily better.

We will revise the explanation to focus on the actual soft-margin optimization behavior rather than presenting receptive-field leakage as a mathematical justification.

Regarding distance consistency, all predicted and target features are L2-normalized, so squared Euclidean distance satisfies

∥e^−e∥22=2(1−cos⁡(e^,e))\|\hat{e}-e\|_2^2=2(1-\cos(\hat{e},e)) Thus, the training alignment distance and inference cosine dissimilarity share the same normalized feature geometry.

AuthorsOpenReview ↗

Clarification of the independent contributions of RA-CL.

We appreciate the reviewer's observation regarding the ablation results and would like to clarify the experimental configurations and the distinct contributions of our two proposed mechanisms.

First, under the same multi-scale Mamba-CNN configuration, replacing Standard CL with RA-CL changes I-AUROC from 0.977 to 0.978, while AUPRO@1% decreases from 0.455 to 0.449. However, this comparison is conducted under an intermediate M-CNN architecture without reciprocal cross-modulation and should not be interpreted as the primary evidence for the independent effectiveness of RA-CL or the performance of our complete architecture.

To independently evaluate RA-CL, we conducted a controlled comparison without our second contribution. Standard CL achieves 0.972/0.455, whereas RA-CL achieves 0.974/0.455 (I-AUROC/AUPRO@1%). This demonstrates that our region-aware formulation provides additional image-level discriminative benefits while maintaining low-FPR localization performance. More importantly, introducing contrastive learning into normal-only cross-modal prediction constitutes the core of our first contribution, while region awareness further enhances this objective. This is also reflected in our staged ablation, where incorporating RA-CL after MS improves both metrics from 0.962/0.448 to 0.974/0.455. Thus, our first contribution encompasses not only region-aware weighting, but also the introduction of discriminative contrastive supervision, further enhanced through mask-conditioned pseudo-anomaly construction and region-aware weighting. Together, these results support the effectiveness of our first contribution as a whole.

AuthorsOpenReview ↗

Clarification of the independent contributions of reciprocal cross-modulation.

Regarding our architectural contribution, we would like to clarify that M-CNN provides the local-global feature extraction foundation, whereas the key innovation lies in the proposed reciprocal Mamba-CNN cross-modulation, rather than the combination of CNN and Mamba itself. Therefore, the M-CNN intermediate configuration should be evaluated together with the subsequent cross-modulation mechanism to understand the contribution of our complete architectural design.

To directly isolate the effect of reciprocal cross-modulation, we fixed RA-CL and the M-CNN feature extraction architecture, and compared configurations with and without CM. Adding reciprocal cross-modulation improves I-AUROC from 0.978 to 0.980 and AUPRO@1% from 0.449 to 0.454. Notably, it recovers most of the localization degradation observed in the intermediate M-CNN configuration, while further improving image-level detection. This provides direct evidence that the benefit comes not merely from combining local CNN and global Mamba representations, but from enabling their adaptive bidirectional interaction for more effective cross-modal alignment.

We will clarify the distinction between these intermediate and controlled ablations in the revised presentation, making the independent contributions of RA-CL and reciprocal cross-modulation, as well as their complementary effects in the complete framework, more explicit.

Reviewer_p32BOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes MambaX-AD, an unsupervised anomaly detection framework integrating RGB images and 3D point clouds. It employs region-aware contrastive learning (RA-CL) to improve anomaly discrimination and reciprocal Mamba-CNN cross-modulation to enhance cross-modal feature mapping, achieving competitive results on two public benchmarks.

Critical strengths
  1. The research problem is practically relevant and meaningful.
  2. The paper provides clear and informative visualizations. Figures 2 and 3 are well-designed and effectively illustrate the proposed method and its architecture.
Critical weaknesses
  1. The proposed method mainly combines existing techniques, offering limited methodological novelty and insufficient insights into the underlying research problem.
  2. Moreover, the Figure 1 does not convincingly demonstrate the existence of the two claimed challenges: weak discriminative sensitivity and the cross-modal alignment bottleneck. Ideally, it should illustrate the failure cases of existing methods followed by the corresponding proposed solutions.
  3. The illustration in Figure 2 should be revised, as Mamba is not equivalent to a State Space Model (SSM). The authors are encouraged to refer to the original Mamba paper for a more accurate representation.
  4. The ablation results in Table 7 show only marginal performance improvements, which are insufficient to convincingly demonstrate the effectiveness of the proposed components.
Questions for authors
  1. Could the authors provide empirical evidence demonstrating the existence of the two claimed challenges, namely weak discriminative sensitivity and the cross-modal alignment bottleneck? How do the proposed modules specifically address these limitations?
  2. Could the authors provide comparisons with alternative architectures, such as Transformer-based models? This would help clarify whether Mamba is essential to the improvements.
8 comments
AuthorsOpenReview ↗

Clarification of the two methodological contributions and their ablation evidence.

We sincerely thank the reviewer for recognizing the practical relevance of our research problem and the clarity of Figures 2 and 3. We appreciate the constructive concerns regarding methodological novelty, the empirical motivation of the two challenges, and the architectural comparisons.

We would like to further clarify the distinct roles of our two proposed contributions and the corresponding experimental evidence.

Our first contribution, Region-Aware Contrastive Learning (RA-CL), introduces contrastive supervision into normal-only cross-modal feature mapping and further enhances it through mask-conditioned pseudo-anomaly construction and region-aware contrastive weighting. The key idea is to jointly exploit clean positive targets and spatially structured pseudo-anomalous negative targets to provide discriminative supervision for cross-modal predictions. This goes beyond standard normal feature alignment and generic feature augmentation.

Our second contribution, reciprocal Mamba-CNN cross-modulation, establishes bidirectional, condition-aware interactions between local CNN features and global Mamba representations. Unlike conventional static, single-mapping frameworks that insufficiently account for the adaptive relationships between local and global information, different modalities, and feature depths, our design enables each branch to generate modulation signals that adaptively regulate the other, with direction- and depth-aware embeddings conditioning these interactions. Rather than simply concatenating or adding outputs, it facilitates adaptive bidirectional feature exchange.

The staged ablation provides evidence for these two contributions.

ConfigurationI-AUROC ↑AUPRO@1% ↑
Baseline0.9540.455
Baseline + MS0.9620.448
+ RA-CL0.9740.455
+ Reciprocal CM (Full)0.9820.456

Adding MS increases I-AUROC to 0.962 but decreases AUPRO@1% to 0.448. After incorporating RA-CL, both metrics improve to 0.974 and 0.455, respectively. Finally, incorporating the second proposed mechanism further improves the results to 0.982/0.456. We therefore attribute the overall improvement to the complementary effects of the proposed discriminative learning and reciprocal interaction mechanisms, rather than to the mere combination of existing feature extractors.

AuthorsOpenReview ↗

Empirical evidence for the two challenges.

We thank the reviewer for suggesting empirical evidence to support the two identified challenges. Although the rebuttal interface does not allow us to upload additional visualizations, representative failure cases are already presented in Fig. 4 of our submission.

For Challenge 1, Fig. 4 (4th column, Foam) shows that CFM fails to adequately highlight the subtle defective region, incorrectly treating part of the anomalous region as normal, reflecting the problem of insufficient discriminative sensitivity leading to missed subtle defects. In contrast, our method successfully identifies this subtle defect, illustrating how RA-CL enhances discriminative sensitivity beyond normal feature alignment.

For Challenge 2, Fig. 4 (2nd, 5th, 8th, and 9th columns, Carrot, Tire, Peach, and Cable G., respectively) shows that CFM and MAFR produce diffuse anomaly responses extending into normal regions surrounding the actual defects, reflecting the problem of insufficient adaptive local-global interaction leading to diffuse anomaly responses and imprecise localization. In contrast, our method achieves more spatially precise localization while suppressing these spurious responses, demonstrating the benefit of reciprocal cross-modulation in facilitating adaptive local-global interactions.

In the revised manuscript, we will further incorporate these representative comparisons into Fig. 1, explicitly connecting the observed failure patterns of existing methods with the two challenges and illustrating how our proposed mechanisms address them.

AuthorsOpenReview ↗

Mamba versus SSM in Figure 2.

We thank the reviewer for pointing out the distinction between Mamba and a generic State Space Model (SSM) operation.

Mamba is a complete selective state-space-based architecture and should not be treated as synonymous with a generic SSM operation. The SSM is an internal component of Mamba, which also includes input-dependent selection and additional projection, convolution, and gating operations. We will revise Figure 2 to distinguish the complete Mamba block from its internal selective SSM component and align the illustration with the original Mamba formulation.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.