acceptodds
Withdrawn from ICLR 2027

Primary-Complementary Bidirectional Enhancement and Query-wise Modality Preference Selection for Visible-Infrared Object Detection

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Existing visible-infrared object detectors commonly fuse heterogeneous features into a single representation for detection. However, this paradigm may weaken modality-specific discriminative cues and fail to explicitly model object-level modality preferences. To address these challenges, we propose BEPS-Net, a query-based visible-infrared object detection framework. Specifically, we introduce Primary-Complementary Bidirectional Enhancement (PCBE), which treats each modality as the primary representation in one enhancement direction and incorporates complementary information from the other. This design maintains distinct visible-primary and infrared-primary representations, helping preserve modality-specific discriminative cues while exploiting cross-modal complementarity. Furthermore, we introduce Query-wise Modality Preference Selection (QMPS), which integrates query initialization from both representations with adaptive modality routing, allowing each object query to dynamically reassess its modality preference at every decoder layer and select informative cues from the two enhanced representations. Extensive experiments demonstrate that BEPS-Net outperforms existing state-of-the-art visible-infrared detectors on multiple benchmarks, while the proposed designs can be effectively extended to visible-infrared semantic segmentation. The code is available at: https://anonymous.4open.science/r/BEPSNet.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Visible-infrared object detection, cross-modal enhancement, modality preference selection

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

This paper proposes the BEPS-Net, a visible-infrared object detection framework. In this method, it designs a Primary-Complementary Bidirectional Enhancement (PCBE) strategy to treat one modality enhancement by the other modality, preserving modality-specific discriminative cues while exploiting cross-modal complementarity. Moreover, it introduces a Query-wise Modality Preference Selection (QMPS) scheme to integrate query initialization from both modality feature representations with modality preference routing. Experiments demonstrate the effectiveness of proposed BEPS-Net.

Critical strengths

There are some Critical Strengths in this paper, including

  1. A visible-infrared object detection framework is proposed to address the potential attenuation of modality-specific discriminative cues.
  2. Primary-Complementary Bidirectional Enhancement scheme is designed to complementarily enhance one modality by the other modality.
  3. Query-wise Modality Preference Selection scheme is designed to decide which modality is prefered prefered in visible-infrared object detection.
  4. This paper is with clear organization structure and completed presentation, and extensive experiments are done to demonstrate the effectiveness of BEPS-Net.
Critical weaknesses

This paper has some Critical Weaknesses:

  1. The main ideas in BEPS-Net, including Primary-Complementary Bidirectional Enhancement and Modality Preference Selection, are not new. For example, the former has been widely explored in DEYOLO (2024), DRPFNet (2026), LDFE (2026), etc. This paper combines existing schemes to construct BEPS-Net framework, with a-little weak scheme novelty.
  2. In the explanation to BEPS-Net, only the mathematical formula of module-level input-ouput is presented. The novelty deepth seems coarse and weak. Moreover, the necessary loss function is not defined.
  3. In experiments, only mAP is used for perforamnce evaluation comparison. More performance metrics could be adopted. Moreover, only the feature responses before and after PCBE are shown. Besides resultant evidenes, more visualization evidences of intermediate process are not reported to compare BEPS-Net with other methods.
Questions for authors

Similar to the "Critical Weaknesses", I only have three concern points, including:

1.The novelty and contribution

The main ideas in BEPS-Net, including Primary-Complementary Bidirectional Enhancement and Modality Preference Selection, are not new. For example, the former has been widely explored in DEYOLO (2024), DRPFNet (2026), LDFE (2026), etc. This paper combines existing schemes to construct BEPS-Net framework, with a-little weak scheme novelty.

2.Weak theory analysis and explanation

In the explanation to BEPS-Net, only the mathematical formula of module-level input-ouput is presented. The novelty deepth seems coarse and weak. Moreover, the necessary loss function is not defined.

3.Performance Metrics and visualization evidences

In experiments, only mAP is used for perforamnce evaluation comparison. More performance metrics could be adopted. Moreover, only the feature responses before and after PCBE are shown. Besides resultant evidenes, more visualization evidences of intermediate process are not reported to compare BEPS-Net with other methods.

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

This paper proposes BEPS-Net for visible-infrared object detection. It introduces Primary-Complementary Bidirectional Enhancement (PCBE) to preserve modality-specific features while incorporating complementary information. It also proposes Query-wise Modality Preference Selection (QMPS) to select different feature branches for object queries. Experiments on several benchmarks show improved detection performance. The method is also extended to semantic segmentation.

Critical strengths

S1. good motivation. The paper considers different modality preferences for different objects. Preserving two modality-enhanced representations is a reasonable approach to RGB-T detection.

S2. Good experimental performance. BEPS-Net achieves competitive results on multiple benchmarks. The authors also provide component ablations, cross-dataset evaluation, and semantic segmentation experiments.

Critical weaknesses

W1. Limited methodological novelty. PCBE mainly combines existing channel attention, spatial attention, and cross-attention modules. Its difference from previous bidirectional fusion methods is not sufficiently clear. QMPS uses hard Top-1 routing, but the paper lacks comparisons with query-wise soft gating. The necessity of these designs is not fully justified.

W2. Insufficient validation of modality preference learning. The router selects between two representations that already contain information from both modalities. The routing statistics only show branch selection frequencies. They do not prove that the router learns modality reliability. Experiments under RGB or infrared degradation are needed to support this claim.

Questions for authors

Q1. Why is hard Top-1 routing necessary? Could the authors compare it with query-wise soft gating under the same settings?

Q2. Does the router actually learn modality reliability? Could the authors evaluate routing decisions under RGB or infrared degradation?

Q3. What is the main technical difference between PCBE and existing bidirectional cross-modal attention methods?

Q4. Figure quality. Figure 2 is poorly presented. The text is too small, and the layout is overly crowded. It is difficult to understand the overall architecture and data flow. Could the authors redraw this figure with a clearer layout and more readable labels?

1 comment

Dear Reviewers,

The authors have posted their rebuttal responses. I would kindly encourage you to review the authors' responses and continue the discussion by providing follow-up comments where appropriate. In particular, please indicate whether the authors' responses have adequately addressed your concerns and update your assessment if necessary.

Thank you for your time and continued efforts!

Your AC

Reviewer_M5GVOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work proposes BEPS-Net to address infrared-visible multimodal object detection. Specifically, the authors introduce Primary-Complementary Bidirectional Enhancement (PCBE) to construct modality-dominant enhanced features and Query-wise Modality Preference Selection (QMPS) to dynamically select modality information for different object queries during decoding. Experiments on multiple detection datasets show competitive performance, and the applicability of the method is further evaluated on semantic segmentation.

Critical strengths
  1. This work starts from the observation that different objects may rely differently on infrared and visible information, and accordingly designs bidirectional feature enhancement and query-level modality selection. The two modules are closely related, resulting in a coherent overall design.

  2. The experiments are relatively comprehensive, with competitive detection performance on LLVIP, FLIR, and M³FD. The paper also includes cross-dataset testing, module ablations, and semantic segmentation experiments, which provide further support for the effectiveness of the method.

Critical weaknesses
  1. The proposed PCBE mainly relies on existing techniques such as channel attention, spatial attention, and cross-attention for bidirectional feature interaction, while the competitive query initialization also shows some similarity to prior methods such as DAMSDet. Although QMPS introduces layer-wise Top-1 routing, its improvement over performing modality selection only in the first layer is relatively limited, making the overall methodological novelty less evident.

  2. Figure 4 mainly presents modality selection ratios across different categories, which is insufficient to show whether QMPS can adapt its modality preference to the actual imaging conditions of different instances within the same category. In addition, the current ablation studies do not clearly separate the gains from preserving dual-branch features and those from the specific enhancement operations.

Questions for authors
  1. The authors are encouraged to further discuss the substantive differences between BEPS-Net and existing query-level multimodal detection methods such as DAMSDet, particularly regarding the novelty of the modality selection mechanism.

  2. Figure 4 mainly presents category-level modality selection ratios. It would be helpful to provide routing analyses for instances of the same category under different illumination or modality degradation conditions to further validate the effectiveness of instance-level modality preference.

  3. In Table 9, layer-wise selection provides only a limited improvement over selection in the first decoder layer. The authors could further analyze how modality selection changes across decoder layers and whether repeated layer-wise updates are necessary.

1 comment

Dear Reviewers,

The authors have posted their rebuttal responses. I would kindly encourage you to review the authors' responses and continue the discussion by providing follow-up comments where appropriate. In particular, please indicate whether the authors' responses have adequately addressed your concerns and update your assessment if necessary.

Thank you for your time and continued efforts!

Your AC

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

Related papers

Discussion (0)

Sign in to comment.