acceptodds
Under review as a conference paper at ICLR 2027

DeCo: A Multimodal Detection Framework based on Density Consensus Learning

Abstract

Multimodal object detection aims to exploit complementary observations from heterogeneous sensors for accurate object localization. However, existing fusion approaches mainly rely on feature-level interaction without explicitly modeling object distribution and spatial organization, limiting their performance in dense scenes. This limitation is particularly evident in remote sensing applications such as optical and SAR ship detection. In this paper, we propose DeCo, a multimodal detection framework based on density consensus learning. DeCo formulates multimodal fusion as consensus estimation over a shared object distribution space, where feature interaction is guided by cross-modal spatial agreement and density consensus, enabling robust multimodal detection under high target density. Specifically, modality-specific encoders transform optical and SAR observations into a shared latent density space, where spatially varying modality reliability is estimated to construct a cross-modal density consensus representation. The learned consensus subsequently regulates multiscale feature fusion by adaptively allocating modality contributions and enhancing spatially consistent object evidence. To supervise density learning without manual density profile design, we introduce a box-aware Bayesian density objective that optimizes instance-level mass assignment through probabilistic ownership modeling. Extensive experiments on paired optical–SAR ship detection benchmarks demonstrate that DeCo achieves robust multimodal detection and accurate object counting, especially in densely packed scenes.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multimodal detection, density consensus learning, feature fusion, box-aware Bayesian

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_UDz3OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes DeCo, a multimodal OPT-SAR detection framework based on density consensus learning. Instead of directly aligning heterogeneous optical and SAR features, DeCo projects both modalities into a shared target-density space, estimates spatially varying modality reliability, and uses the resulting density consensus to guide multiscale feature fusion. A box-aware Bayesian density objective is further introduced to supervise instance-level density mass without predefined Gaussian density targets. Experiments are conducted on DenseShip6K and MOS-Ship.

Critical strengths

The idea of establishing cross-modal agreement in a target-distribution space rather than directly matching OPT and SAR appearance features is intuitive and well motivated. The ablations are reasonably complete. In particular, the paper shows that density estimation alone does not explain the gain; performance improves when the learned density consensus is explicitly used to guide feature fusion.

Critical weaknesses

1.The main dense-scene evaluation relies heavily on synthetic SAR data. DenseShip6K generates SAR images from optical images, so the two modalities are more tightly coupled and better aligned than real OPT-SAR observations. This weakens the evidence for the claimed advantage under real heterogeneous dense scenes.

2.The robustness of density consensus to cross-modal misalignment is unclear. The method relies on spatial agreement between OPT and SAR features, but no sensitivity analysis to registration errors is provided. This is important for practical OPT-SAR fusion.

Questions for authors

1.Can the authors evaluate DeCo on a real, densely populated OPT-SAR dataset, or provide a dense-scene subset analysis on MOS-Ship?

2.How sensitive is DeCo to spatial misalignment between optical and SAR observations?

3.Can the authors provide a stronger same-backbone comparison between ordinary multimodal fusion and DeCo on the real MOS-Ship dataset?

5 comments
AuthorsOpenReview ↗

Response to Reviewer UDz3 — Comments 1–2

Weakness 1/ Question 1: Evaluation on Real Dense OPT-SAR Scenes

We thank the reviewer for raising this concern regarding the synthesized SAR data in DenseShip6K. We agree that synthesized SAR counterparts provide a controlled dense-scene setting but cannot fully reproduce the sensing characteristics and cross-modal heterogeneity of real OPT-SAR observations.

To address this limitation, we conducted additional experiments on QXS-Ship, a manually annotated subset constructed from the real paired OPT-SAR dataset QXS-SAROPT. It contains 620 paired 256×256 images with 8,907 annotated ship instances, including positive and negative samples. All images are manually annotated with horizontal bounding boxes using the optical observations as the primary reference.

We use the best checkpoint trained and validated on MOS-Ship to evaluate QXS-Ship without additional training or fine-tuning. The density distribution is summarized below.

Density levelNumber of imagesShip instances
Negative1440
Low-density260492
Medium-density971,187
High-density1197,228
Total6208,907

The high-density subset contains 119 images and 7,228 ship instances, accounting for 81.1% of all annotated ships. As reported in Table 14, DeCo achieves 79.7% AP50 and 40.7% AP50:95 on the complete test set, with an MAE of 3.25 and an RMSE of 6.15. On the high-density subset, it achieves 83.5% AP50 and 42.6% AP50:95, with 91.6% Precision and 76.1% Recall.

These results provide additional evidence of DeCo's effectiveness on real OPT-SAR observations, including highly concentrated ship distributions. The dataset description, density distribution, qualitative visualizations, and density-stratified evaluation are provided in Appendix C.3.


Weakness 2/ Question 2: Sensitivity to Cross-Modal Spatial Misalignment

We thank the reviewer for pointing out the importance of spatial misalignment in practical OPT-SAR fusion. We conducted controlled translation experiments on QXS-Ship by shifting SAR images relative to their corresponding optical images by 2, 4, 8, and 16 pixels in four cardinal directions, while keeping the optical images and ground-truth annotations unchanged. We use the MOS-Ship checkpoint without training or fine-tuning on QXS-Ship. For each nonzero displacement magnitude, we report the mean and standard deviation across the four directions; the 0-pixel row is the original aligned baseline.

SAR shiftAP50AP50:95MAERMSE
0 px79.7040.703.256.15
2 px79.68 ± 0.0440.73 ± 0.043.26 ± 0.046.19 ± 0.09
4 px79.65 ± 0.1740.78 ± 0.083.27 ± 0.036.21 ± 0.13
8 px79.58 ± 0.2940.73 ± 0.083.31 ± 0.086.27 ± 0.30
16 px79.05 ± 0.3640.30 ± 0.073.42 ± 0.126.46 ± 0.36

DeCo maintains relatively stable performance under small spatial offsets. As the displacement increases, performance gradually degrades, with more noticeable changes in recall and counting errors. At 16 pixels, AP50 decreases from 79.70% to 79.05%, while AP50:95 decreases from 40.70% to 40.30%. MAE and RMSE increase from 3.25 and 6.15 to 3.42 and 6.46, respectively.

These results suggest that DeCo exhibits a degree of tolerance to moderate cross-modal translation errors, while accurate spatial correspondence remains important for effective density-consensus-guided fusion. The experimental protocol and results are provided in Appendix C.3.2 and Table 15.

AuthorsOpenReview ↗

Response to Reviewer UDz3 — Comment 3

Question 3. Same-Backbone Comparison on Real MOS-Ship Data

We thank the reviewer for this valuable suggestion. We have conducted additional same-backbone experiments on the real MOS-Ship dataset, as reported in Appendix C.2.1.

All variants use the same backbone and detection head, with only the input modality and fusion strategy varied. We compare OPT-only and SAR-only baselines with four multimodal fusion strategies: feature concatenation, element-wise addition, cross-attention, and generic spatial-adaptive fusion (GSAF).

Fusion strategyPrecisionRecallAP50AP50:95
OPT-only80.372.880.553.4
SAR-only73.861.071.642.3
Concatenation76.565.476.249.0
Element-wise Addition75.264.974.746.8
Cross-Attention77.668.278.150.1
Generic Spatial-Adaptive Fusion79.674.181.053.2
DeCo81.775.581.954.7

Conventional fusion strategies do not consistently improve detection performance over the OPT-only baseline. In contrast, DeCo achieves the best performance across all four metrics. Compared with GSAF, which incorporates conventional spatial modulation and adaptive modality weighting, DeCo improves Precision, Recall, AP50, and AP50:95 by 2.1, 1.4, 0.9, and 1.5 percentage points, respectively. These controlled comparisons provide further evidence that the performance improvements of DeCo are not simply attributable to the use of additional modalities or conventional feature-level fusion operations.

Dear Reviewer UDz3,

We sincerely thank you for your insightful comments and constructive suggestions, which have been very helpful in improving our work. We have revised the paper, and the changes are indicated in blue text within the manuscript.

We hope that our responses have adequately addressed the concerns you raised in your review, and we would greatly appreciate your further feedback. Please let us know if there is anything further we can clarify.

Best regards, The Authors

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

This paper proposes the DeCo, a multimodal object detection framework with density consensus learning. It formulates traditional multimodal fusion as consensus estimation over a shared object distribution space of cross-modal spatial agreement and density consensus. In this scheme, modality-specific encoders will transform optical and SAR observations into a shared latent density space. Moreover, the learned consensus subsequently regulates multiscale feature fusion by adaptively allocating modality contributions and enhancing spatially consistent object evidences. This paper also introduces a box-aware Bayesian density objective to optimize instance-level mass assignment. Extensitve experiments on two paired OPT-SAR ship datasets (MOS-Ship and DenseShip6K) are performed to prove the effectiveness of proposed DeCo framework.

Critical strengths

There are several Critical Strengths could be observed in this paper.

  1. This paper is with clear organzation structure and presentation.
  2. DeCo, a multimodal detection framework, is proposed. It formulates traditional cross-modal fusion as a consensus estimation over a shared target-distribution space.
  3. In Deco, a density-guided interaction mechanism is designed to learn modality-specific density evidence and spatially varying modality reliability. Moreover, a box-aware Bayesian density objective is designed to constrain instance-wise density mass through probabilistic ownership assignment.
  4. Extensive experiments on two paired OPT-SAR ship datasets (MOS-Ship and DenseShip6K) are performed.
Critical weaknesses

Besides some Critical Strengths, this paper is also with some Critical Weaknesses, especially in experiments. They are briefly listed as follows.

  1. In the explanation to DeCo, the key part is unclear what the Density is. It is pure convlution results, Kernel Density or other else? In Eqn. 5, only dilated conv and residual shortcut‌ are seen.
  2. In perforamnce comparison of Table 1, no 2026-published methods are compared. I wonder if these comapred methods have truly contained current SOTA ones.
  3. In terms of performance evaluation metrics, besides APs, MAE (Eqn. 41) and RMSE (Eqn. 42) are adopted. As an object detection task, some classic and popular metrics, e.g., Precision and Recall, are not used in Table 1. They are used only in Table 3. Totally, the performance metrics used in in differnt tables are not always consistent. Moreover, essentially, seen from their computation equations, both MAE and RMSE are also highly-homologous.
  4. The model complexity comparison of different methods are unclear, e.g., parameter scales and FLOPs.
Questions for authors

I have several concerns for this paper. Please further explain and clarify them in rebuttal.

1.The explanation to key Density Estimation

In the explanation to DeCo, the key part is unclear what the Density is. It is pure convlution results, Kernel Density or other else? In Eqn. 5, only dilated conv and residual shortcut‌ are seen.

2.Compared methods

In perforamnce comparison of Table 1, no 2026-published methods are compared. I wonder if these comapred methods have truly contained current SOTA ones.

3.Inconsistent or density performance metrics

In terms of performance evaluation metrics, besides APs, MAE (Eqn. 41) and RMSE (Eqn. 42) are adopted. As an object detection task, some classic and popular metrics, e.g., Precision and Recall, are not used in Table 1. They are used only in Table 3. Totally, the performance metrics used in in differnt tables are not always consistent. Moreover, essentially, seen from their computation rules, both MAE and RMSE are also highly-homologous.

4.Model complexity

The model complexity comparison of different methods are unclear, e.g., parameter scales and FLOPs.

5 comments
AuthorsOpenReview ↗

Response to Reviewer feuv — Comments 1–2


We are very grateful to the reviewer for the insightful comments and constructive suggestions, which have been very helpful for improving our work. In response to the feedback from all three reviewers, we have revised the paper, and the changes are indicated in blue text within the manuscript. In the following, we provide a point-by-point response to your concerns.

Weakness 1/ Question 1: Clarification of the Density Representation

We thank the reviewer for pointing out the lack of clarity regarding the density representation in DeCo. We have revised Sections 3.2 and 3.4 and the explanation following Eq. (5) to distinguish the intermediate density representation from an explicit density map.

In DeCo, density refers to a learned spatial representation of target-distribution evidence, rather than a kernel density estimate (KDE) or a normalized probability density function. Specifically, Eq. (5) employs parallel dilated context operators to aggregate target-related spatial information at different receptive fields, while the residual connection preserves the original multiscale features. The resulting EmE_m is therefore a density-aware latent feature, not an explicit pixel-wise density map.

These modality-specific features are projected into a shared latent density space to obtain ZoZ_o and ZsZ_s, from which the spatially weighted consensus ZcZ_c is constructed. The explicit density map D^\hat D is then generated by decoding ZcZ_c and supervised through the box-aware Bayesian mass-assignment objective.

We further clarify that the Gaussian-shaped likelihood in Eq. (13) is used only to calculate the probabilistic ownership of each pixel among individual objects and the background. It does not generate predefined Gaussian density targets or perform kernel density estimation.


Weakness 2/ Question 2: Comparison with Recent Methods Published in 2026

We thank the reviewer for this suggestion. We have included MHFNet (Zhang et al., 2026), a recently published framework for OPT-SAR ship detection, in the main benchmark comparison. We adopt its official implementation for evaluation.

DatasetMethodAP50AP50:95
DenseShip6KMHFNet83.456.4
DenseShip6KDeCo85.058.5
MOS-ShipMHFNet80.754.1
MOS-ShipDeCo81.954.7

DeCo achieves higher detection accuracy than MHFNet on both datasets. We have updated the compared-method description in Section 4.1, the main results in Table 1, and the corresponding discussion in Section 4.2. MHFNet is also included in the additional comparisons in the revised appendix.

AuthorsOpenReview ↗

Response to Reviewer feuv — Comments 3–4

Weakness 3 / Question 3: Evaluation Metrics

We thank the reviewer for raising this concern. In the revised manuscript, AP50 and AP50:95 are retained as the primary detection metrics in the main benchmark comparison. AP summarizes the precision-recall curve at a given IoU threshold, while AP50 and AP50:95 evaluate detection performance at IoU 0.50 and across IoU thresholds from 0.50 to 0.95, respectively. Precision and Recall are additionally reported in selected ablation and sensitivity analyses (Tables 3 and 15), controlled fusion comparisons (Table 8), and density-stratified evaluations (Table 14).

Regarding the counting metrics, we retain both MAE and RMSE because they capture different characteristics of counting errors. MAE measures the average magnitude of absolute deviations, whereas RMSE assigns greater penalties to large errors and is more sensitive to severe counting failures. This distinction is particularly relevant in dense scenes, where a small number of substantial counting errors may significantly affect overall performance.


Weakness 4 / Question 4: Model Complexity

We thank the reviewer for this suggestion. We have added a model complexity analysis in Appendix C.2.2, comparing parameter counts and FLOPs of representative methods, including TridentNet, Sparse R-CNN, DINO, ICAFusion, DPAL-R, EI2Det, MHFNet, and DeCo.

As reported in the newly added Table 9, DeCo contains 38.54M parameters and requires 176.70G FLOPs. Its computational cost is lower than that of DPAL-R and EI2Det, while achieving higher detection accuracy. Compared with MHFNet, DeCo introduces additional computational overhead. However, DeCo achieves consistently better performance than MHFNet on all datasets across all evaluation metrics, with clear improvements. These results characterize the trade-off between detection accuracy and computational complexity in DeCo.

Dear Reviewer feuv,

We hope that our responses have adequately addressed the concerns you raised in your review, and we would greatly appreciate your further feedback. Please let us know if there is anything further we can clarify.

Best regards, The Authors

Reviewer_iv4iOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses multimodal object detection in remote sensing imagery, focusing on dense ship detection using complementary optical (OPT) and synthetic aperture radar (SAR) observations. The authors propose DeCo, a density consensus learning framework consisting of three main components: (1) Cross-modal Density Consensus Estimation, which projects OPT and SAR features into a shared latent density space and learns spatially adaptive modality weights to construct a consensus representation; (2) Density-guided Multiscale Fusion, which uses the learned consensus to adaptively fuse multimodal features and enhance target-related responses; and (3) Box-aware Bayesian Density Supervision, which encourages instance-level density mass assignment without predefined pixel-wise Gaussian density targets. Experiments on DenseShip6K and MOS-Ship demonstrate improved detection and counting performance compared with existing methods, particularly in densely populated scenes and under degraded optical conditions.

Critical strengths
  • The combination of optical and SAR images is well-motivated, as the two modalities provide complementary information. Leveraging their complementary strengths for robust ship detection is a reasonable and meaningful research direction.
  • The proposed method achieves superior detection and counting performance compared with existing methods on both DenseShip6K and MOS-Ship. The extensive ablation studies further support the effectiveness of the proposed components.
Critical weaknesses
  • The proposed density consensus is essentially a weighted summation of two modality-specific latent features (Eq. 7), which closely resembles conventional adaptive feature fusion. Moreover, the modality-specific representations are not explicitly supervised as density distributions, and the method does not directly measure cross-modal agreement in object distributions. Therefore, it is unclear what fundamentally distinguishes the proposed density consensus from ordinary density-supervised feature fusion.
  • The spatially adaptive modality weights are essential to the proposed framework, as they determine which modality contributes more reliable information. However, the paper only reports averaged reliability weights under clean and degraded optical conditions, without visualizing their spatial distributions or validating whether they correctly identify unreliable regions.
  • Density-guided object detection has already been explored in prior works, including Li et al. (2020a), Duan et al. (2021), and Zhao et al. (2026), which are also cited in the paper. Similarly, adaptive modality weighting and feature modulation are established techniques in multimodal fusion. The proposed framework mainly combines these existing ideas, without clearly demonstrating a sufficiently novel mechanism or explaining its fundamental differences from prior density-guided detection approaches.
  • The proposed structural consistency loss encourages OPT and SAR to exhibit similar normalized spatial feature-energy distributions. However, the two modalities are expected to provide complementary information, particularly when one modality is degraded. For example, when ships are obscured by clouds in optical images but remain visible in SAR, their corresponding feature responses may naturally differ. Penalizing such differences could potentially suppress useful modality-specific information. Although the authors investigate the sensitivity to the loss weight, its effect on preserving complementary information under modality-specific degradation remains insufficiently analyzed.
  • As shown in Tables 3 and 5, introducing density supervision or density consensus alone does not improve detection performance, while the main improvements arise from the proposed consensus-guided feature fusion. However, it remains unclear whether these gains originate from the claimed density consensus mechanism or simply from additional spatial modulation and adaptive feature weighting.
Questions for authors

Please see the Weaknesses

7 comments
AuthorsOpenReview ↗

Response to Reviewer iv4i — Comment 1


We are very grateful to the reviewer for the insightful comments and constructive suggestions, which have been very helpful for improving our work. In response to the feedback from all three reviewers, we have revised the paper, and the changes are indicated in blue text within the manuscript. In the following, we provide a point-by-point response to your concerns.

Weakness 1: Distinction of the Density Consensus Mechanism

We thank the reviewer for this insightful comment. We agree that the original manuscript did not sufficiently clarify the distinction between the proposed density consensus and conventional adaptive feature fusion. We have revised Sections 2, 3.2, 3.3, and 3.4, and added controlled experiments in Appendix C.2.1.

1. Construction of the consensus representation.

We acknowledge that Eq. (7) is formulated as a weighted aggregation. The distinction lies in how its inputs and weights are constructed and used. Specifically, the modality-specific representations ZoZ_o and ZsZ_s are first projected into a shared latent target-distribution space. Their concatenated representations are then used to jointly estimate spatially varying modality reliability. Thus, ZcZ_c aggregates target-related evidence according to the estimated local reliability of each modality, rather than directly combining heterogeneous backbone features.

2. Density supervision.

We clarify that ZoZ_o and ZsZ_s are modality-specific latent representations encoding target-distribution evidence, not independently predicted density maps. The only explicit density prediction in DeCo is the consensus-derived map D^\hat{D}, decoded from ZcZ_c.

The box-aware Bayesian mass-assignment objective and the global count constraint supervise D^\hat{D}. Their gradients propagate through the consensus representation and modality-reliability gating to both modality-specific branches. In addition, the cross-modal structural consistency loss regularizes the normalized spatial energy distributions derived from ZoZ_o and ZsZ_s, without requiring identical feature responses across modalities.

3. Difference from ordinary density-supervised fusion.

Density consensus directly mediates cross-modal interaction and participates in the detection pathway. The modality reliability maps regulate the contribution of each sensor at each spatial location, while the consensus representation provides spatial guidance for multiscale feature fusion.

Our progressive ablation in Table 3 includes ordinary multimodal fusion with an auxiliary density task but without density consensus. This configuration does not improve detection over ordinary fusion. Moreover, introducing density consensus without using it to guide detection features does not produce the final performance gains. Table 5 further verifies the complementary contributions of modality-reliability guidance and consensus-aware spatial modulation.

We have also introduced a generic spatial-adaptive fusion baseline (GSAF), which incorporates conventional spatial modulation and adaptive modality weighting without density consensus estimation or density-based supervision. Its controlled comparison with DeCo is reported in Table 8.

Together, these clarifications and experiments distinguish the proposed mechanism from ordinary density-supervised fusion and conventional spatial-adaptive feature fusion.

AuthorsOpenReview ↗

Response to Reviewer iv4i — Comment 2

Weakness 2: Spatial Visualization of Modality Reliability

We thank the reviewer for this insightful comment. We agree that averaged modality reliability weights alone are insufficient to demonstrate their spatial adaptivity. We have therefore extended Appendix C.2.3 with spatial visualizations and region-wise quantitative evaluation.

First, we visualize the optical reliability maps WoW_o under clean and cloud-degraded optical conditions, together with the corresponding changes in SAR reliability ΔWs\Delta W_s, using paired samples from MOS-Ship. The same trained DeCo model is evaluated using the clean and degraded OPT observations, while the corresponding SAR images and ground-truth annotations remain unchanged. As illustrated in the newly added Fig. 9, the optical reliability maps exhibit spatially varying responses to cloud degradation, while the corresponding ΔWs\Delta W_s maps reveal increased relative contributions from SAR in affected regions.

Second, we divide the ground-truth ship regions into cloud-affected and clean regions according to the cloud degradation mask, and calculate the average SAR reliability within each region. The results are reported in Table 12.

RegionClean Wˉs\bar W_sCloud-degraded Wˉs\bar W_sIncrease
Clean0.19340.3037+0.1103
Cloud-affected0.24450.4226+0.1781

The increase in SAR reliability is larger in cloud-affected regions (+0.1781) than in clean regions (+0.1103). These results provide quantitative evidence that the learned reliability mechanism responds more strongly to regions where optical observations are degraded.

These findings complement the aggregate ship-region statistics in Table 11, where the average optical reliability decreases from 0.7580 to 0.5832, while SAR reliability increases from 0.2420 to 0.4168 under cloud degradation.

Together, Fig. 9 and Tables 11–12 demonstrate that the learned modality reliability maps capture spatially varying changes in observation quality and adaptively redistribute the relative contributions of OPT and SAR.

AuthorsOpenReview ↗

Response to Reviewer iv4i — Comment 3

Weakness 3: Novelty Compared with Existing Density-Guided Detection Methods

We thank the reviewer for pointing out the related studies on density-guided object detection and adaptive multimodal fusion. We have revised the Related Work section and further clarified the methodological distinctions in Sections 3.2 and 3.3.

Specifically, DMNet (Li et al., 2020a) and CDMNet (Duan et al., 2021) utilize density maps to identify object-rich regions and guide adaptive image cropping or cluster generation. More recently, DRMNet (Zhao et al., 2026) employs density maps as spatial priors for dense-region focusing and feature interaction. These studies demonstrate the effectiveness of density information in improving object localization and feature representation within aerial detection pipelines.

The distinction of DeCo lies in the role and utilization of density representations. Unlike the above methods, which primarily exploit density as an intra-modal spatial prior for region selection or feature refinement, DeCo formulates multimodal fusion as consensus estimation over a shared latent target-distribution space. Heterogeneous OPT and SAR features are mapped into this common space, where spatially varying modality reliability is jointly estimated to construct a cross-modal density consensus. The resulting consensus provides modality-reliability-based feature allocation and consensus-aware spatial modulation, both of which are incorporated into multiscale detection feature fusion.

Furthermore, DeCo introduces a box-aware Bayesian mass-assignment objective to supervise the decoded consensus density at the instance level, without requiring predefined pixel-wise Gaussian density targets. Cross-modal structural consistency provides a soft constraint on the spatial organization of modality-specific representations while preserving their heterogeneous responses.

The experiments further support this distinction. As shown in Table 3, introducing an auxiliary density task into conventional multimodal fusion does not improve detection performance, whereas the complete density-consensus-guided framework achieves substantially better results. In addition, the GSAF baseline, which incorporates conventional spatial modulation and adaptive modality weighting without density consensus, is outperformed by DeCo on MOS-Ship.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.