acceptodds
Under review as a conference paper at ICLR 2027

BASFR-Net: Learning Stagewise Residual Interactions for RGB-Infrared Oriented Object Detection

Abstract

RGB-infrared oriented object detection in aerial remote sensing must exploit complementary appearance and thermal cues while preserving the spatial detail of small, arbitrarily oriented objects. Existing multimodal object detection methods often optimize a particular cross-modal fusion mechanism, although the role of interaction changes across feature extraction, multiscale aggregation, and prediction. We present BASFR-Net, a compact detector that formulates cross-modal fusion as stagewise residual interaction. It preserves base representations while progressively calibrating modality responses, reconciling semantic and spatial information, and routing local evidence to prediction. BASFR-Net achieves 84.30% and 76.80% [email protected] on DroneVehicle and VEDAI, outperforming a feature-concatenation baseline by 2.00 and 7.60 points, respectively. On DroneVehicle, it reaches 69.64% [email protected]:0.95 with 5.06M parameters and 13.38 GFLOPs. Ablations show consistent gains from each interaction stage. These results establish stagewise residual interaction as an effective and compact design principle for RGB-infrared oriented object detection.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
multimodal object detection, RGB-infrared oriented object detection, cross-modal fusion, stagewise residual interaction, aerial remote sensing

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_FMeyOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes BASFR-Net, a stagewise residual interaction framework for RGB-infrared oriented object detection in aerial remote sensing. The method consists of three core modules: BAFG for cross-modal response calibration at the backbone stage, SFAB for semantic-detail feature reconstruction at the neck stage, and CRAH for local feature routing at the prediction stage. Each module is designed to address the interaction requirements of a different stage, while residual connections preserve the original feature representations. Experiments on the DroneVehicle and VEDAI datasets demonstrate substantial performance improvements.

Critical strengths

1.Effective module design. The ablation studies show that each module independently improves detection performance, while their combination yields further gains, demonstrating the effectiveness and complementarity of the proposed modules.

2.Clear dataset visualizations. The qualitative results illustrate challenging scenarios involving densely packed vehicles, missed detections, and classification confusion. BASFR-Net reduces missed detections and classification errors, broadly consistent with the improvements observed in the quantitative results.

Critical weaknesses
  1. The experiments are conducted on only two datasets, making it difficult to establish the generalization capability of the proposed method. Additional evaluation is needed.
  2. The comparison with existing state-of-the-art methods is insufficient. Results for larger model variants, such as YOLO11s and YOLO11m, are also missing.
  3. The meanings of the symbols used in the loss functions are not adequately explained.
  4. The figures and writing have substantial shortcomings. (a) The overall framework diagram is overly simplified, and some symbols and labels are unclear. The core designs, methodological contributions, and residual paths are not clearly presented, reducing readability. (b) The contribution section claims that explicit paths are retained and updated at each stage, but it does not sufficiently explain what is updated or how the stages differ. The contributions are insufficiently connected to the methodological details. (c) The related-work section does not establish sufficiently clear connections between existing fusion and alignment methods and BASFR-Net’s stage-specific residual interactions for backbone response calibration, neck aggregation, and prediction-stage routing.
  5. The motivation and novelty of the method are unclear. Although the paper discusses multimodal fusion and small-object detection, the method section primarily describes the proposed modules without clearly explaining how they address the identified challenges. The claimed stagewise modeling strategy is mainly manifested as separate module designs for the backbone, neck, and head, without sufficiently demonstrating why different modules are specifically suited to their respective stages.
  6. For SFAB, the paper should clarify the differences between semantic and detail inputs, how these features are obtained, and the practical motivation for combining them. For CRAH, the five spatial priors require more detailed explanations, as does the allocation of routing weights across different feature scales. The illustration of how modality cues guide local evidence aggregation before oriented prediction is also overly simplistic.
  7. The ablation studies are insufficient. Some hyperparameter choices lack supporting experiments and quantitative justification.
Questions for authors
  1. Generalization: Additional datasets and cross-dataset generalization experiments are recommended.
  2. SOTA comparisons: More comparisons with state-of-the-art methods and additional necessary ablation studies are recommended.
  3. Methodological novelty: The motivation for stagewise interaction should be clarified, and its necessity should be validated through experiments comparing different module placements and combinations.
  4. Module mechanisms: The sources and roles of SFAB’s input features should be explained, along with the five spatial priors in CRAH and its multiscale routing mechanism.
  5. Figures and presentation: The overall architecture and module diagrams should be improved to clarify residual paths and symbol definitions. The explanations of mathematical notation and the paper’s formatting should also be revised.
Reviewer_hLrjOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

BASFR-Net targets oriented vehicle detection on paired RGB and infrared aerial images. Two YOLO11n streams feed three interaction stages: BAFG calibrates each modality with an affine proposal from the other and a disagreement weight, SFAB reassembles semantic and detail branches with content-aware kernels, and CRAH routes fused features over a 3x3 neighbourhood at P3 and P4. Each stage keeps the concatenated inputs as an additive identity path. Experiments use DroneVehicle and VEDAI, reporting 84.30 and 76.80 [email protected] with 5.06M parameters and 13.38 GFLOPs.

Critical strengths
  1. The three stages are not copies of one fusion block: each has a different role and adds a bounded update to a retained path, so the ablation can separate them instead of measuring only the stack.

  2. The same detector is tested on two datasets where the stronger modality flips, with per-class AP for both. On DroneVehicle the sequential ablation moves 82.30, 83.70, 84.30, the paper's most convincing evidence that the stages add up.

Critical weaknesses
  1. The premise for infrared is illumination, stated on page 1, yet no experiment conditions on illumination: night, low-light, fog and adverse return nothing. The only qualitative evidence is Figure 6 on page 9, two unlabelled scenes, with no VEDAI example and no failure case. Without a day and night split, 84.30 and 76.80 cannot be attributed to the lighting regime.

  2. Every number comes from a single run, while the claims rest on deltas of 0.49 to 0.90 points. No seed, standard deviation or significance test appears. Section 4.1 states that the component study uses the same DroneVehicle test set throughout, and the ablation exists only there, so the 7.60-point VEDAI gain has no component attribution.

  3. The full model adds 1.19M parameters and 3.56 GFLOPs with no capacity-matched control, so part of the 2.00-point gain may come from size rather than the interaction. Table 1 gives no parameters, GFLOPs or input resolution for the published rows, the VEDAI split is never stated, and Equation (10) omits the three loss weights.

Questions for authors
  1. Table 1 cites published VEDAI scores rather than re-running them, which only holds if your setup matches theirs. What split, class set and input size did you use?

  2. Figure 6 has two daytime scenes, so the gain may come mostly from daylight. Please report DroneVehicle results for day and night.

  3. The 0.60 to 0.90 point deltas come from single runs and may sit within noise. Could you repeat Table 2 with three seeds and standard deviations?

  4. The model adds 1.19M parameters, so size may explain part of the gain. Please add a capacity-matched control such as a wider YOLO11n.

  5. Table 2 and Appendix Table 6 ablate modules on DroneVehicle only, so nothing shows which module produces the VEDAI gain. Please repeat Table 2 on VEDAI.

  6. In Equation (2), D is both the map and channel dropout, so the second term is ambiguous. Could you clarify whether s in (4) is the gate s_i?

  7. Without the three loss weight values in Equation (10), the numbers cannot be reproduced. Please report them and the VEDAI setup.

  8. VEDAI is reported only at IoU 0.5, where small objects are easy to match. Could you report [email protected]:0.95 there as you do for DroneVehicle?

Reviewer_m6HcOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents BASFR-Net, a dual-stream YOLO11n-OBB framework for RGB–infrared oriented object detection in aerial images. The network introduces three modules: BAFG for cross-modal feature calibration, SFAB for semantic and spatial feature refinement, and CRAH for feature aggregation before prediction.Experiments are conducted on DroneVehicle and VEDAI. BASFR-Net achieves 84.30% and 76.80% [email protected] on the two datasets, respectively, outperforming the authors' feature-concatenation baseline. The ablation results also indicate that each proposed module contributes to the final performance.

Critical strengths
  1. The overall framework is reasonably designed. The three modules are placed at different stages of the network and address different aspects of multimodal feature fusion. The architecture is clearly described, and the implementation appears relatively straightforward.

  2. The experimental results are encouraging. BASFR-Net improves the baseline performance on both DroneVehicle and VEDAI, with particularly noticeable gains on VEDAI. The ablation experiments further show that the individual modules are useful and that combining them leads to additional improvements.

Critical weaknesses
  1. The methodological novelty is somewhat limited. The proposed modules rely heavily on existing ideas, including affine feature modulation, CARAFE, FreqFusion, and local feature aggregation. The main contribution seems to lie in how these techniques are combined and arranged within the detection framework. I am not fully convinced that the proposed stagewise residual interaction represents a sufficiently distinct technical contribution. In addition, the full model requires about 31% more parameters and 36% more computation than the baseline. It would be useful to determine whether similar improvements could be obtained simply by increasing the capacity of the baseline or using simpler fusion modules.

  2. The experimental evaluation needs to be more convincing.Although the reported results are competitive, the improvements over the strongest existing methods are relatively small (0.49% on DroneVehicle and 0.52% on VEDAI). It is also unclear whether all compared methods follow the same experimental settings. No results from multiple random seeds are provided, and the exact data split for VEDAI is not clearly specified. Furthermore, the ablation experiments are conducted only on DroneVehicle, and some combinations of the proposed modules are missing. These issues make it difficult to judge the robustness of the improvements and the actual contribution of the three-stage design.

Questions for authors
  1. Have the authors compared BASFR-Net with a baseline of similar parameter count and computational cost? A comparison with a larger YOLO11n-OBB baseline or simpler fusion modules would help clarify where the performance improvements come from.

  2. Could the authors clarify the data splits and evaluation protocols, particularly for VEDAI? Were the results of the compared methods obtained under identical settings? Reporting the mean and standard deviation over multiple runs would also make the comparison more reliable.

  3. Why are the BAFG+CRAH and SFAB+CRAH combinations missing from the ablation study? It would also be helpful to include ablation results on VEDAI to verify whether the three modules consistently complement each other across datasets.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.