acceptodds
Withdrawn from ICLR 2027

BRIDGE:Boosting Representation Fidelity via Shape-Informed Distance-field Geometric Encoding for Multimodal Object Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Multimodal object detection seeks to exploit complementary information across heterogeneous sensing modalities for robust object perception. Different modalities reveal distinct structural and geometric cues of the same object through different imaging principles. However, heterogeneous imaging mechanisms complicate cross-modal geometric interaction, while complex object geometries remain difficult to capture with shapes parameterized by only a few parameters. In this paper, we introduce BRIDGE, a novel geometric representation framework for object detection that moves beyond predefined shapes toward continuous geometry modeling. Specifically, we observe that Signed Distance Fields (SDF) map heterogeneous modalities into a shared continuous space while preserving fine-grained geometry. We therefore introduce Continuous Geometry Encoding (CGE) to represent objects beyond predefined parametric shapes. We further propose Fourier-Conditioned Geometry Fusion (FCGF) to preserve informative geometric cues while exploiting cross-modal complementarity through frequency-conditioned residual refinement. Finally, we dive into SDF optimization during training and introduce Gradient-Modulated Loss (GML). GML suppresses gradients from well-fit regions while strengthening corrections for severe geometric deviations. Mathematical analysis and experiments show that SDF provides bounded and spatially distributed optimization signals for stable multimodal geometric learning. BRIDGE achieves 85.0% mAP on DroneVehicle, outperforming existing multimodal detectors and demonstrating the effectiveness of continuous geometry modeling. The code and models will be available on GitHub: xxxxx to facilitate future work.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multimodal Object Detection, Oriented Object Detection, Geometric Encoding, Signed Distance Fields, Cross-Modal Fusion

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_R2PoOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes BRIDGE, a geometric modeling framework for multimodal oriented object detection, consisting of three main components: Continuous Geometry Encoding (CGE), Fourier-Conditioned Geometry Fusion (FCGF), and Gradient Modulation Loss (GML). CGE uses signed distance fields (SDFs) to provide a unified representation of oriented bounding boxes and object contours. FCGF estimates modality reliability based on image frequency statistics and performs geometric fusion through boundary gating and cross-modal consistency gating. GML adjusts the gradient contributions during optimization according to the magnitude of geometric residuals.

Critical strengths

A key strength of this work is its attempt to improve multimodal detection from three complementary perspectives: geometric representation, cross-modal fusion, and loss design. The authors provide corresponding theoretical analyses and component-level ablation studies. Experimental results demonstrate that the proposed framework achieves competitive detection performance on the DroneVehicle dataset. In addition, SAM3 is used only to generate contour supervision offline, introducing no additional inference overhead. Overall, the method has practical engineering value, and the experiments provide preliminary support for several of its design choices.

Critical weaknesses

The advantages of continuous geometric representations receive some empirical support, but the methodological novelty and relative advantages remain insufficiently justified.

The authors adopt SDFs as the core geometric representation and provide theoretical analysis to explain how they encode geometric information at individual spatial locations. However, converting object boundaries into SDFs is an established geometric representation technique. The main contribution of this component therefore lies in applying SDFs to multimodal oriented object detection and optimizing the model with spatially distributed supervision.

In Table 3, CGE achieves an mAP50 of 72.5%, outperforming GWD at 71.3%. This result suggests that the proposed representation offers certain advantages under the current experimental setting. However, it is insufficient to establish the general superiority of SDFs over alternative geometric representations. Furthermore, performance increases to 75.8% when SAM3-based contour supervision is introduced. Since this improvement also involves finer-grained supervisory information, the gain cannot be attributed solely to the SDF representation itself.

The core assumption underlying FCGF lacks direct validation, and the paper has not demonstrated that it can reliably identify the modality with better geometric quality.

FCGF uses mid- and high-frequency Fourier statistics to estimate the structural quality of RGB and infrared images, selects a dominant modality accordingly, and incorporates geometric corrections from the auxiliary modality through spatial gating. This design relies on a critical assumption: image frequency statistics can effectively reflect the reliability of the geometric information provided by each modality.

However, frequency responses are not directly equivalent to geometric quality. High-frequency components may arise from clear object boundaries and fine details, but they may also originate from noise or other forms of interference. The authors themselves acknowledge in the supplementary material that the proposed Fourier quality score is only a proxy for structural quality and does not directly measure the error of the predicted SDF.

The existing experiments primarily demonstrate changes in the modality-weight distribution after score calibration, as well as performance improvements resulting from boundary gating and consistency gating. These results do not directly establish that the frequency score can accurately identify the modality with better geometric quality, nor do they adequately isolate the contribution of the frequency-based reliability estimation itself.

For this core design, the authors should provide evidence establishing at least one of the following two capabilities. First, under realistic noise conditions or explicitly controlled image degradations, the method should reliably identify the more informative modality and improve detection performance. Second, even when input image quality is controlled or denoising is applied, the method should still identify the modality containing more accurate geometric information and achieve a verifiable improvement in geometric quality through fusion. The current experiments do not sufficiently establish either line of evidence.

Therefore, the problem with FCGF is not merely that its reliability metric relies heavily on empirical heuristics. More fundamentally, there is insufficient direct experimental evidence connecting its core assumption to actual geometric quality.

The motivation for GML is not sufficiently established, and the current mathematical analysis does not demonstrate that its optimization advantages are generalizable.

GML adjusts gradient magnitudes according to the magnitude of SDF residuals. For the final choice of (p=1.5), the gradient magnitude increases linearly with the residual when (r leq 1), while growing proportionally to (sqrt{r}) when (r>1). Compared with the (L_1) loss, this design reduces the per-sample gradient contribution of small residuals while increasing the relative influence of larger residuals. Compared with the squared loss, its gradient grows more slowly in the large-residual regime.

However, this design raises two concerns.

First, the choice of the piecewise threshold (r=1) is primarily motivated by the distribution of normalized residuals. Although the supplementary material reports relevant statistics, it does not adequately explain why this threshold is a reasonable and stable choice. Nor do the authors establish its necessity through systematic threshold-sensitivity experiments.

Second, the mathematical analysis mainly demonstrates properties such as continuity, differentiability, and sublinear gradient growth as the residual increases. These properties do not directly establish that a large number of small residuals will not dominate the overall optimization process, nor do they guarantee an appropriate balance between emphasizing large residuals and maintaining optimization stability. The actual parameter gradients also depend on the network Jacobian, the number of samples, and the directions of the individual sample gradients.

Although the existing ablation studies show that (p=1.5) outperforms several alternative settings under the reported experimental configurations, this only demonstrates the effectiveness of that choice in the current experiments. It is insufficient to establish the general validity of the underlying optimization rationale. GML therefore appears to be a specific loss design with some empirical support, while its independent methodological novelty remains insufficiently substantiated.1. The advantages of continuous geometric representations receive some empirical support, but the methodological novelty and relative advantages remain insufficiently justified.

The authors adopt SDFs as the core geometric representation and provide theoretical analysis to explain how they encode geometric information at individual spatial locations. However, converting object boundaries into SDFs is an established geometric representation technique. The main contribution of this component therefore lies in applying SDFs to multimodal oriented object detection and optimizing the model with spatially distributed supervision.

In Table 3, CGE achieves an mAP50 of 72.5%, outperforming GWD at 71.3%. This result suggests that the proposed representation offers certain advantages under the current experimental setting. However, it is insufficient to establish the general superiority of SDFs over alternative geometric representations. Furthermore, performance increases to 75.8% when SAM3-based contour supervision is introduced. Since this improvement also involves finer-grained supervisory information, the gain cannot be attributed solely to the SDF representation itself.

Consequently, although the paper provides preliminary comparative evidence, it does not adequately disentangle the respective contributions of the geometric representation, additional contour supervision, and associated optimization strategies. The current results support the effectiveness of the overall design, but they do not sufficiently establish a clear and generalizable representational advantage.

Questions for authors

This paper has practical engineering value, and its three components address distinct aspects of the detection pipeline: geometric representation, cross-modal fusion, and optimization. The existing experiments support the effectiveness of some design choices under specific datasets and configurations. Therefore, it would be inappropriate to dismiss the method as a mere collection of components with no practical effect.

Nevertheless, from the perspective of methodological novelty, SDFs, Fourier frequency statistics, gated residual fusion, and piecewise gradient modulation are all built upon existing techniques. The primary contribution of this work lies in combining these techniques and adapting them to a specific task, rather than introducing a new principle supported by sufficiently strong theoretical or empirical evidence.

More importantly, the paper does not adequately validate the core reliability assumption underlying FCGF, nor does it sufficiently establish a causal connection between the specific design choices in GML and the resulting performance improvements. The existing ablation studies demonstrate that certain components are effective under the current configurations, but they do not establish that these components are indispensable or provide stable generalization advantages.

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

This paper proposes BRIDGE, a novel geometric representation framework for multimodal object detection. By utilizing the Signed Distance Fields (SDF) for unified representation, it validates the effectiveness of continuous geometric modeling on the DroneVehicle dataset.

Critical strengths

This paper introduces Continuous Geometry Encoding (CGE), Fourier-Conditioned Geometry Fusion (FCGF), and Gradient-Modulated Loss (GML) within a unified multimodal object detection framework. This integration provides a solution to address geometric bias and heterogeneous modal reliability in multimodal object detection. In particular, it offers a reasonable mechanism for constraining auxiliary geometric corrections to ensure boundary proximity and cross-modal consistency. BRIDGE achieved an mAP50 of 85.0% on the DroneVehicle dataset with only 18.96 million parameters, outperforming existing multimodal detectors. Extensive ablation experiments were conducted, demonstrating the effectiveness of each component’s design. These results validate the effectiveness of continuous geometric modeling.

Critical weaknesses

I am interested in the overall innovativeness of the paper. The paper reports that the SDF representation outperforms GWD and further achieves performance improvements by utilizing contours generated by SAM3; however, it is currently not possible to fully distinguish whether the gains stem from the SDF representation itself or from the additional contour supervision. I recommend that the authors conduct a more rigorous comparison under identical supervision conditions. I am concerned about the reproducibility of the methods and the fairness of the comparisons. The main experiments in the paper used a single random seed and did not report standard deviations or confidence intervals from multiple runs; therefore, it is difficult to determine whether the 0.6 percentage point improvement on DroneVehicle is stable. Furthermore, the main results table primarily summarizes performance reported in previous work, and it remains to be clarified whether the training configurations and evaluation protocols were fully consistent across different methods.

Questions for authors

Can the authors report results across multiple random seeds and clarify whether baseline comparisons use consistent training configurations and evaluation protocols? Can the authors isolate the benefits of SDF representation from those of SAM3-derived contour supervision under identical supervision conditions?

Reviewer_PQ62OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces BRIDGE, a continuous geometry framework for multimodal oriented object detection. It proposes Continuous Geometry Encoding (CGE), which represents object geometry through signed distance fields in a shared continuous space, and Fourier-Conditioned Geometry Fusion (FCGF), which uses frequency-domain cues and spatial gates to guide cross-modal geometric refinement. Gradient-Modulated Loss (GML) adjusts optimization according to geometric residual magnitude, reducing the influence of small errors while strengthening corrections for larger deviations. Experiments demonstrate competitive performance on DroneVehicle. Additional evaluations on RGB–SAR imagery support the applicability across modalities.

Critical strengths

BRIDGE represents geometry through shared SDFs, explicitly encoding boundary distances and extending beyond OBBs to irregular contours. In addition, this paper provides detailed mathematical analyses of SDF perturbation stability, spatially distributed gradients, and the GML response. CGE, FCGF, and GML work together to unify geometric representation, cross-modal interaction, and optimization within a framework. The reported results show competitive detection accuracy and parameter count, with additional evaluations across sensing modalities and detection representations.

Critical weaknesses

1.Please clarify how the features F1,F2 are converted into the geometric fields S1,S2 in Figure 2. It is unclear whether these fields are predicted directly from features or constructed from predicted bounding boxes, as Equation (3) only defines the box-to-SDF transformation.

2.The gradient perturbation bound in Section 3.2, Eq. (8), is derived for a squared objective in Appendix A.4, whereas Appendix E.4 specifies L1 and GML supervision. Please clarify the scope of this guarantee and whether analogous bounds hold for the implemented objectives, particularly L1 near zero residual.

3.Please clarify whether the complete BRIDGE model in Table 1 uses the same experimental configuration as the ablations in Table 2. The complete model is reported at 85.0% mAP50, whereas the two-component configurations in Table 2 reach 74.3% and 74.0%. Please include the full CGE+FCGF+GML configuration under the ablation protocol and explain how it relates to the 85.0% result.

4.Section 4.1 states that both mAP50 and mAP50:95 are reported in line 377, but the result tables only present mAP50. Please provide the missing mAP50:95 results or revise the statement accordingly.

5.Qualitative detection comparisons between the baseline and BRIDGE under matched settings, together with predicted contours before and after FCGF fusion, would help assess how geometric fusion affects localization.

Questions for authors

The authors are suggeted to clarify the theoretical foundations and implementation details, ensure consistency between the main results and ablation studies, provide more comprehensive evaluation results, and include additional qualitative comparisons to better demonstrate the effectiveness of the proposed method.

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

Related papers

Discussion (0)

Sign in to comment.