Pith. sign in

REVIEW 4 major objections 4 minor 39 references

Deep Multimodal Fusion Detection through Spatial Mask and Channel Fusion

T0 review · 4 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper claims that splitting visible-infrared fusion into a training-time spatial exchange and an inference-time channel competition yields the best reported detection results across five benchmarks.

desk verdict Plausible fusion design and clean ablations, but the reported SOTA margins are internally inconsistent—particularly on VEDAI and DroneVehicle—so the empirical claims need correction before they are trusted. read the letter →

arxiv 2608.02092 v1 pith:FYIV2URU submitted 2026-08-03 cs.CV cs.MM

Guandi Wang , Ming Li , Yunsen Xing , Junle Liu This is my paper · ORCID
classification cs.CVcs.MM
keywords RGB-infraredobjectdetectionmultimodalfusionchannelcompetitionsemanticmaskexchangespatial-channelattentionmodalityasymmetryaerialentmax-basedproxy
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Visible and infrared images are physically different signals, so a single fusion operator asked both to integrate complementary features and to shut off a degraded modality faces an optimization conflict. The paper's answer is to split the problem: during training, a Semantic Mask Exchange swaps semantically divergent spatial blocks between the two branches so each backbone learns without relying on fixed modality labels; at inference, a Learnable Channel Competition makes a hard per-channel choice of which modality's feature to keep. The authors report that this two-part framework, ADCR, outperforms earlier fusion methods on all five datasets tested, with the largest margins on DroneVehicle, an aerial benchmark where the modalities are both severely asymmetric and spatially misaligned. The body of the paper reports top results in its comparison tables, while the abstract describes the outcome more modestly as competitive.

What carries the argument

The load-bearing object is ADCR's shared Spatial-Channel Attention (CSA), which produces per-channel and per-pixel confidence masks for both modalities under identical weights and normalization. The authors claim this makes the visible and infrared confidence scores comparable, so the discrepancy between them is an information-content measure. On that foundation, SME and LCC are the two levers: SME computes a Semantic Cosine Distance between block-aggregated spatial masks to choose where to swap features during training; LCC converts the channel discrepancy into a hard per-channel routing decision, using the Adaptive Boundary Proxy, an Entmax-inspired convex interpolation, only to give the d

What would settle it

Run the identical ADCR pipeline but replace the shared CSA with two independent attention projections, one per modality, keeping all other hyperparameters fixed. If the DroneVehicle margin over the next-best method, or the LCC ablation gain on FLIR, shrinks substantially, then the comparability of the attention weights, not the competition mechanism, is the source of the reported improvement.

Watch

Extended reading notes

Core claim

At the core of the paper is the claim that the failure of existing RGB-infrared fusion in asymmetric conditions is an optimization conflict, not a capacity problem. The same function cannot softly average two healthy modalities and also hard-gate a corrupted one, so the paper decouples these duties. SME is a training-only augmentation that computes a block-level exchange score from the product of a spatial-attention gap and a semantic cosine distance, selects top-K blocks, and Bernoulli-samples which blocks to swap between branches; this keeps the backbone from over-specializing to one modality. LCC is an inference-time channel router: using a shared Spatial-Channel Attention, it takes the p

Load-bearing premise

The load-bearing premise is that the shared Spatial-Channel Attention makes visible and infrared confidence scores directly comparable, so the per-channel discrepancy reflects information content rather than photometric differences; if that fails, LCC's hard channel choices could be driven by arbitrary sensor bias.

Editorial extensions

If this is right

  • If ADCR's central claim is right, dual-backbone detection can be made more reliable under modality imbalance without extra loss functions: SME adds no parameters and no inference cost.
  • Hard per-channel routing with sparse gradients can beat soft weighted averaging when one modality is degraded; the largest reported margins are on aerial and misaligned data.
  • Fusion design splits into two independent subproblems—where to complement (spatial) and how to aggregate (channel)—which are optimized separately, potentially simplifying future fusion architectures.
  • The framework is plug-and-play for existing dual-branch detectors; the paper reports roughly 1.09M added parameters and 21 GFLOPs at the deepest feature stage.
  • The learnable stiffness parameter for each channel adapts to dataset asymmetry, producing more decisive suppression on FLIR than on VEDAI, which suggests the router learns a dataset-specific policy rather than a fixed one.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • We infer the same decoupling principle should transfer to other heterogeneous sensor pairs, such as depth-camera, LiDAR-camera, or polarization, and to video where modality quality fluctuates over time; the paper does not test these settings.
  • The paper itself notes in its visualization section that SME sometimes redistributes modality-specific noise between branches rather than suppressing it; we infer that this trade-off is dataset-dependent and worth measuring explicitly.
  • The sensitivity peak at p=0.6 for the Bernoulli sampling probability hints at an optimal mixing dose; we infer that an adaptive, per-block probability governed by the exchange score could improve on the fixed schedule, though the paper does not run that experiment.
  • We infer that the shared-CSA comparability premise is the part to stress-test first, because the paper's ablation does not isolate it; a control with non-shared attention projections would clarify whether LCC's gains come from the selection mechanism or from the shared projection.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes Attention-Driven Complementarity Resampling (ADCR), a training-inference decoupled fusion framework for visible-infrared object detection. During training, a Semantic Mask Exchange (SME) module swaps spatial blocks between modality branches based on a semantic cosine distance computed from shared spatial-channel attention weights. During inference, a Learnable Channel Competition (LCC) module selects, per channel, the modality with higher attention-derived confidence, using a straight-through estimator with an adaptive boundary proxy for differentiability. Experiments are reported on LLVIP, M3FD, FLIR, VEDAI, and DroneVehicle, with ablations on FLIR and sensitivity analysis of the Bernoulli sampling probability. The central claim is that ADCR outperforms existing fusion methods on all five benchmarks, with the largest margins on DroneVehicle.

Significance. The conceptual separation of training-time spatial augmentation from inference-time channel arbitration is a reasonable and potentially useful direction, and the reported ablations indicate that both SME and LCC contribute positively. The efficiency analysis showing that LCC adds only 0.001M parameters is valuable. The paper also states that source code is provided in the supplementary material, which is a strength for reproducibility. However, the empirical claims are undermined by internal inconsistencies in the reported tables, and the key theoretical assumption about cross-modal comparability of attention weights is asserted without direct evidence. If corrected and verified, the framework could be a solid contribution, but in its current form the headline results are not reliable.

major comments (4)
  1. [§4.1, Table 1] The claim that on VEDAI 'ADCR surpasses all evaluated multi-modal fusion methods' is contradicted by the table itself. multimodal DINO is reported at 83.6 mAP50 / 53.8 mAP95, while ADCR (CSPDarknet53-v8) reaches 82.7 / 49.5 and ADCR (v5) reaches 81.9 / 50.2. Both ADCR variants are worse than DINO on both metrics. Please correct the text or the table and re-assess the 'all five benchmarks' statement.
  2. [§4.1, Table 2] The DroneVehicle margins '4.9% and 8.5 points' mix two backbone variants. The sentence states 'ADCR achieves 84.1% mAP50 and 64.5 mAP95', but 84.1 is from the v8 row and 64.5 is from the v5 row. Against the best fusion competitor CMAFF (82.0/57.6), the margins are at most 2.4/6.9 (v5) or 2.1/6.7 (v8). Against Fusion-Mamba v8 (79.2/56.0), the margins are 5.2/8.5 (v5) or 4.9/8.3 (v8). Please report margins within a single configuration and avoid cherry-picking across rows.
  3. [§3, Eqs. (3)-(4)] The Bernoulli sampling in Eq. (4) uses s_{i,j} as a probability, but s_{i,j} is defined as a product of a normalized gap (bounded by 1) and a cosine-distance term. If the block features are not guaranteed nonnegative, the cosine distance can exceed 1, making s_{i,j} > 1 and thus an invalid Bernoulli parameter. Even under nonnegative features, the text's description of 'orthogonality' implying distance 1 omits the opposite-vector case (distance 2). Please clarify the bounds and any clamping/normalization used in the actual implementation.
  4. [§3, CSA comparability] The load-bearing premise that shared convolutional weights make W_vis and W_ir 'purely content-driven' comparable is asserted but not experimentally isolated. Since LCC's hard channel selection depends entirely on the sign of Δc, please provide evidence for this assumption: e.g., compare a shared-CSA variant against separate CSAs, or show that Δc correlates with measured per-modality image degradation. Without such evidence, the mechanism underlying LCC remains a plausible heuristic rather than a validated principle.
minor comments (4)
  1. [§4.2, Table 3 vs text] The text says removing LCC and using ConcatConv 'decreases the performance to 77.7% in mAP50', but Table 3 reports 77.5. Please harmonize the numbers.
  2. [§4.1, Implementation] The phrase 'untrained CSPdarknet' is ambiguous. If the backbone is trained from scratch, say so; if it is pretrained, specify the initialization. Also, no random seeds or error bars are reported; given that some reported gains are below one point (e.g., LLVIP mAP95, M3FD mAP50 vs Fusion-Mamba v8), a single-run comparison is insufficient to support the claimed advantages.
  3. [§4.3, Fig. 4] The sensitivity analysis plots would benefit from error bars or repeated runs, especially because the text acknowledges 'minor localized fluctuations' at adjacent values of p. The conclusion that the peak at p=0.6 is robust is hard to evaluate from single curves.
  4. [General] Some references are incomplete or inconsistently formatted (e.g., [29], [35]); please check the reference list against the publisher's style.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation; the ADCR pipeline is a learned routing mechanism evaluated against external benchmarks.

full rationale

The paper's derivation chain is not circular. The proposed modules (SME and LCC) are defined directly from the shared CSA attention weights W, which are themselves trained end-to-end with the detection loss. LCC routes each channel by the sign of Δc = W_vis,ch − W_ir,ch; this is a learned arbitration rule, not a fitted quantity that is later reported as a prediction. The channel weights are not fit to the reported benchmark numbers and then renamed as a 'prediction' — they are internal parameters of the model, and the central empirical claim (ADCR improves mAP on five benchmarks) is validated against external baselines that are not derived from the paper's own equations. No load-bearing step reduces by construction to its own input: SME's exchange scores are computed from the same attention maps, but that is an architectural design choice (attention-guided augmentation), not a logical circularity. The paper invokes no uniqueness theorem and contains no self-citations by the present authors; the cited prior work (CutMix, Entmax, STE, etc.) provides independent methodological components. The internal inconsistencies in the reported VEDAI and DroneVehicle comparisons (e.g., Table 1 showing multimodal DINO at 83.6 mAP50 vs ADCR at 82.7 mAP50 on VEDAI, and the 4.9/8.5 margins mixing v8 and v5 configurations) are correctness/consistency concerns about the empirical claims, not circularity in the derivation. The assumptions about CSA making attention weights cross-modally comparable are premises, not self-referential reductions. Overall, the derivation is self-contained and externally falsifiable; no circular step is present.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The method's effectiveness hinges on three unproven premises: that shared-parameter attention scores are cross-modally comparable, that attention magnitude tracks information content, and that spatial swapping during training helps rather than hurts. No new physical entities are introduced.

free parameters (5)
  • Bernoulli sampling probability p = 0.6
    Selected after observing the peak on FLIR and M3FD in Fig. 4; this is post-hoc tuning on the datasets used for the final numbers.
  • Spatial partition number N = 10
    Fixed heuristic in Section 4: 'The spatial partition number N and top-k selection ratio in SME are fixed to 10 and 50% respectively throughout all experiments.'
  • top-K selection ratio = 50%
    Fixed heuristic; chosen to 'cover a wide range of candidate swap regions' (Section 4).
  • Logit scale s = unspecified
    Equation (7) calls s a 'predefined logit scale' but no value is given; required for re-implementation.
  • alpha_c interval [1.01, 2.0] = interval [1.01, 2.0]
    The learnable interpolation stiffness is constrained to this range by design (Section 3.2).
assumptions (3)
  • domain assumption Shared CSA parameters make W_vis and W_ir cross-modally comparable; the discrepancy is purely content-driven.
    Section 3, paragraph after Eq. (1): 'the discrepancy between W_vis and W_ir reflects purely content-driven divergence.' LCC's Δc depends on this.
  • domain assumption Attention weight W is a proxy for information density, so regions with low W are task-irrelevant or corrupted.
    Section 3, first paragraph: 'F' = F ⊙ W... making it a natural proxy for information density.'
  • domain assumption Semantic Mask Exchange during training increases downstream robustness without harming the learned representations.
    Section 3.1; the paper hypothesizes this mechanism, and the ablation on FLIR supports it, but it is not proven in general.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Deep Multimodal Fusion Detection through Spatial Mask and Channel Fusion." pith.science (2026). https://pith.science/paper/FYIV2URU

@misc{pith2026260802092,
  author       = {Pith},
  title        = {Pith review of: Deep Multimodal Fusion Detection through Spatial Mask and Channel Fusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FYIV2URU}},
  note         = {Machine review of arXiv:2608.02092}
}
read the original abstract

Deep multimodal fusion for object detection has demonstrated good performance through mining modal characteristics. However, existing feature-level fusion methods mainly weigh between two modalities and unify them in a unified representation space. This can lead to overfitting or over-specialization of the statistical properties of a single modality within a dual-backbone architecture. This paper proposes an Attention-Driven Complementarity Resampling framework for robust improvement of cross-modality object detection. Based on a shared channel spatial attention mechanism, we first introduce the semantic mask exchange to actively mix the boundaries of the modalities during the training phase, forcing the backbone network to learn generalized features without relying on fixed modal labels. Then we propose a learnable channel competition to sample and aggregate features in a channel-wise and learnable way. Our experiments on multiple datasets demonstrate that the proposed method is effective and yields competitive results among existing state-of-the-art approaches. The source code is provided in the supplementary material.

Figures

Figures reproduced from arXiv: 2608.02092 by the authors.

Figure 1
Figure 1. Pipeline of our proposed method. The dashed arrows indicate control flow and the solid arrows indicate the direction of network feature transmission. We select a dual￾backbone feature extraction network as the backbone, and ADCR modules are embedded in three different stages of the network. The two branches do not share their parameters; the channel spatial attention is shared for two modalities. The CSA module empl… view at source ↗
Figure 2
Figure 2. Semantic Mask Exchange Mechanism. The exchange score (N ×N)is calculated by Semantic Cosine Distance (SSD), and is subsequently sparsified through a top-K selec￾tion. The exchange is operated during the training stage, and we set a Bernoulli sampling to regularize the distribution of exchanged features. It utilizes continuous exchange scores si, j as activation probabilities, ensuring that the fusion boundary remain… view at source ↗
Figure 3
Figure 3. Learnable Channel Competition Mechanism. Previous spatially mixed features undergo per-channel arbitration based on the inter-modal confidence discrepancy ∆c. The channel competition branch is controlled by an indicator function I(∆c > 0) that executes a discrete decision. To overcome the non-differentiability of the indicator function, an Adap￾tive Boundary Proxy is constructed using a convex interpolation of a sof… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Sensitivity Analysis of the Bernoulli Sampling Probability. The consistent per￾formance peak observed at the intermediate value (p = 0.6) robustly validates the necessity of our proposed semantics-guided feature circulation during training [PITH_FULL_IMAGE:figures/ful…
Figure 5
Figure 5. Figure 5: , For each scene, we present a block-level exchange score map, the original RGB and infrared(IR) image pairs, and the feature maps before and after SME is applied [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Per channel α deviation from initialization. (a) is the result of VEDAI, (b) is the result of FLIR. α −αinit is the offset of learnable parameter α [PITH_FULL_IMAGE:figures/full_fig_p013_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 3 linked inside Pith

  1. [1]

    Transfusion: Robust lidar-camera fusion for 3d object detection with transformers

    Xuyang Bai, Zeyu Hu, Xinge Zhu, Qingqiu Huang, Yilun Chen, Hongbo Fu, and Chiew-Lan Tai. Transfusion: Robust lidar-camera fusion for 3d object detection with transformers. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1090–1099, 2022

    2022

  2. [2]

    Estimating or propagat- ing gradients through stochastic neurons for conditional computation.arXiv preprint arXiv:1308.3432, 2013

    Yoshua Bengio, Nicholas Léonard, and Aaron Courville. Estimating or propagat- ing gradients through stochastic neurons for conditional computation.arXiv preprint arXiv:1308.3432, 2013

    arXiv 2013

  3. [3]

    Yolov4: Optimal speed and accuracy of object detection.arXiv preprint arXiv:2004.10934, 2020

    Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao. Yolov4: Optimal speed and accuracy of object detection.arXiv preprint arXiv:2004.10934, 2020

    arXiv 2004

  4. [4]

    End-to-end object detection with transformers

    Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kir- illov, and Sergey Zagoruyko. End-to-end object detection with transformers. InEuro- pean conference on computer vision, pages 213–229. Springer, 2020

    2020

  5. [5]

    Weakly misalignment-free adaptive feature alignment for uavs-based multi- modal object detection

    Chen Chen, Jiahao Qi, Xingyue Liu, Kangcheng Bin, Ruigang Fu, Xikun Hu, and Ping Zhong. Weakly misalignment-free adaptive feature alignment for uavs-based multi- modal object detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26836–26845, 2024

    2024

  6. [6]

    Slbaf-net: Super-lightweight bimodal adaptive fusion network for uav detection in low recognition environment.Multimedia Tools and Applications, 82(30):47773–47792, 2023

    Xiaolong Cheng, Keke Geng, Ziwei Wang, Jinhu Wang, Yuxiao Sun, and Pengbo Ding. Slbaf-net: Super-lightweight bimodal adaptive fusion network for uav detection in low recognition environment.Multimedia Tools and Applications, 82(30):47773–47792, 2023. W ANG: MULTIMODAL FUSION DETECTION VIA SPA TIAL MASK AND CHANNEL FUSION15

    2023

  7. [7]

    Fusion-mamba for cross-modality object detection.IEEE Transactions on Multimedia, 2025

    Wenhao Dong, Haodong Zhu, Shaohui Lin, Xiaoyan Luo, Yunhang Shen, Guodong Guo, and Baochang Zhang. Fusion-mamba for cross-modality object detection.IEEE Transactions on Multimedia, 2025

    2025

  8. [8]

    Cross-modality fusion transformer for multispectral object detection.arXiv preprint arXiv:2111.00273, 2021

    Qingyun Fang, Dapeng Han, and Zhaokui Wang. Cross-modality fusion transformer for multispectral object detection.arXiv preprint arXiv:2111.00273, 2021

    arXiv 2021

Show all 39 references
  1. [9]

    Ob- ject detection in thermal spectrum for advanced driver-assistance systems (adas).Ieee Access, 9:156465–156481, 2021

    Muhammad Ali Farooq, Peter Corcoran, Cosmin Rotariu, and Waseem Shariff. Ob- ject detection in thermal spectrum for advanced driver-assistance systems (adas).Ieee Access, 9:156465–156481, 2021

    2021

  2. [10]

    Lraf-net: Long-range attention fusion network for visible–infrared object detection.IEEE Transactions on Neural Networks and Learning Systems, 35 (10):13232–13245, 2023

    Haolong Fu, Shixun Wang, Puhong Duan, Changyan Xiao, Renwei Dian, Shutao Li, and Zhiyong Li. Lraf-net: Long-range attention fusion network for visible–infrared object detection.IEEE Transactions on Neural Networks and Learning Systems, 35 (10):13232–13245, 2023

    2023

  3. [11]

    Multi- spectral pedestrian detection: Benchmark dataset and baseline

    Soonmin Hwang, Jaesik Park, Namil Kim, Yukyung Choi, and In So Kweon. Multi- spectral pedestrian detection: Benchmark dataset and baseline. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 1037–1045, 2015

    2015

  4. [12]

    A review on methods and applications in multimodal deep learning

    Summaira Jabeen, Xi Li, Muhammad Shoib Amin, Omar Bourahla, Songyuan Li, and Abdul Jabbar. A review on methods and applications in multimodal deep learning. ACM Trans. Multimedia Comput. Commun. Appl., 19(2s), February 2023. ISSN 1551-

    2023

  5. [13]

    Mul- tispectral object detection enhanced by cross-modal information complementary and cosine similarity channel resampling modules

    Junbo Jang, Chanyeong Park, Heegwang Kim, Jiyoon Lee, and Joonki Paik. Mul- tispectral object detection enhanced by cross-modal information complementary and cosine similarity channel resampling modules. In2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WAC...

    2025

  6. [14]

    Llvip: A visible- infrared paired dataset for low-light vision

    Xinyu Jia, Chuang Zhu, Minzhen Li, Wenqi Tang, and Wenli Zhou. Llvip: A visible- infrared paired dataset for low-light vision. InProceedings of the IEEE/CVF interna- tional conference on computer vision, pages 3496–3504, 2021

    2021

  7. [15]

    Chenchen Jiang, Huazhong Ren, Hong Yang, Hongtao Huo, Pengfei Zhu, Zhaoyuan Yao, Jing Li, Min Sun, and Shihao Yang. M2fnet: Multi-modal fusion network for object detection from visible and thermal infrared images.International Journal of Applied Earth Observation and Geoinform...

    2024

  8. [16]

    Ultralytics yolo, January 2023

    Glenn Jocher, Jing Qiu, and Ayush Chaurasia. Ultralytics yolo, January 2023. URL https://github.com/ultralytics/ultralytics. AGPL-3.0 license

    2023

  9. [17]

    ultralytics/YOLOv5: v6.0, 2021

    Glenn Jocher et al. ultralytics/YOLOv5: v6.0, 2021

    2021

  10. [18]

    Densefuse: A fusion approach to infrared and visible images

    Hui Li and Xiao-Jun Wu. Densefuse: A fusion approach to infrared and visible images. IEEE transactions on image processing, 28(5):2614–2623, 2018

    2018

  11. [19]

    Crossfuse: A novel cross attention mechanism based infrared and visible image fusion approach.Information Fusion, 103:102147, 2024

    Hui Li and Xiao-Jun Wu. Crossfuse: A novel cross attention mechanism based infrared and visible image fusion approach.Information Fusion, 103:102147, 2024. 16W ANG: MULTIMODAL FUSION DETECTION VIA SPA TIAL MASK AND CHANNEL FUSION

    2024

  12. [20]

    Chunmian Lin, Daxin Tian, Xuting Duan, Jianshan Zhou, Dezong Zhao, and Dongpu Cao. 3d-dfm: Anchor-free multimodal 3-d object detection with dynamic fusion mod- ule for autonomous driving.IEEE Transactions on Neural Networks and Learning Systems, 34(12):10812–10822, 2022

    2022

  13. [21]

    Target-aware dual adversarial learning and a multi-scenario multi- modality benchmark to fuse infrared and visible for object detection

    Jinyuan Liu, Xin Fan, Zhanbo Huang, Guanyao Wu, Risheng Liu, Wei Zhong, and Zhongxuan Luo. Target-aware dual adversarial learning and a multi-scenario multi- modality benchmark to fuse infrared and visible for object detection. InProceedings of the IEEE/CVF conference on compu...

    2022

  14. [22]

    Infrared and visible image fusion methods and applications: A survey.Information fusion, 45:153–178, 2019

    Jiayi Ma, Yong Ma, and Chang Li. Infrared and visible image fusion methods and applications: A survey.Information fusion, 45:153–178, 2019

    2019

  15. [23]

    Swinfu- sion: Cross-domain long-range learning for general image fusion via swin transformer

    Jiayi Ma, Linfeng Tang, Fan Fan, Jun Huang, Xiaoguang Mei, and Yong Ma. Swinfu- sion: Cross-domain long-range learning for general image fusion via swin transformer. IEEE/CAA Journal of Automatica Sinica, 9(7):1200–1217, 2022

    2022

  16. [24]

    Sparse sequence-to-sequence mod- els

    Ben Peters, Vlad Niculae, and André FT Martins. Sparse sequence-to-sequence mod- els. InProceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 1504–1519, 2019

    2019

  17. [25]

    Cross-modality attentive feature fusion for object detection in multispectral remote sensing imagery.Pattern Recognition, 130:108786, 2022

    Fang Qingyun and Wang Zhaokui. Cross-modality attentive feature fusion for object detection in multispectral remote sensing imagery.Pattern Recognition, 130:108786, 2022

    2022

  18. [26]

    Vehicle detection in aerial imagery: A small target detection benchmark.Journal of Visual Communication and Image Rep- resentation, 34:187–203, 2016

    Sebastien Razakarivony and Frederic Jurie. Vehicle detection in aerial imagery: A small target detection benchmark.Journal of Visual Communication and Image Rep- resentation, 34:187–203, 2016

    2016

  19. [27]

    Icafusion: It- erative cross-attention guided feature fusion for multispectral object detection.Pattern Recognition, 145:109913, 2024

    Jifeng Shen, Yifei Chen, Yue Liu, Xin Zuo, Heng Fan, and Wankou Yang. Icafusion: It- erative cross-attention guided feature fusion for multispectral object detection.Pattern Recognition, 145:109913, 2024

    2024

  20. [28]

    Dropout: a simple way to prevent neural networks from overfitting

    Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research, 15(1):1929–1958, 2014

    1929

  21. [29]

    Adaptive multimodal feature fusion with fre- quency domain gate for remote sensing object detection.Remote Sensing Letters, 15 (2):133–144, 2024

    Xu Sun, Yinhui Yu, and Qing Cheng. Adaptive multimodal feature fusion with fre- quency domain gate for remote sensing object detection.Remote Sensing Letters, 15 (2):133–144, 2024

    2024

  22. [30]

    Drone-based rgb-infrared cross- modality vehicle detection via uncertainty-aware learning.IEEE Transactions on Cir- cuits and Systems for Video Technology, 32(10):6700–6713, 2022

    Yiming Sun, Bing Cao, Pengfei Zhu, and Qinghua Hu. Drone-based rgb-infrared cross- modality vehicle detection via uncertainty-aware learning.IEEE Transactions on Cir- cuits and Systems for Video Technology, 32(10):6700–6713, 2022

    2022

  23. [31]

    Deep multimodal fusion by channel exchanging.Advances in neural infor- mation processing systems, 33:4835–4845, 2020

    Yikai Wang, Wenbing Huang, Fuchun Sun, Tingyang Xu, Yu Rong, and Junzhou Huang. Deep multimodal fusion by channel exchanging.Advances in neural infor- mation processing systems, 33:4835–4845, 2020

    2020

  24. [32]

    Cdc-yolofusion: Leveraging cross-scale dynamic convolution fusion for visible-infrared object detec- tion.IEEE Transactions on Intelligent V ehicles, 2024

    Zian Wang, Xianghui Liao, Jin Yuan, You Yao, and Zhiyong Li. Cdc-yolofusion: Leveraging cross-scale dynamic convolution fusion for visible-infrared object detec- tion.IEEE Transactions on Intelligent V ehicles, 2024. W ANG: MULTIMODAL FUSION DETECTION VIA SPA TIAL MASK AND CHA...

    2024

  25. [33]

    If-usod: Multimodal information fusion interactive feature enhancement architecture for underwater salient object detection

    Genji Yuan, Jintao Song, and Jinjiang Li. If-usod: Multimodal information fusion interactive feature enhancement architecture for underwater salient object detection. Information Fusion, 117:102806, 2025

    2025

  26. [34]

    Translation, scale and rotation: cross- modal alignment meets rgb-infrared vehicle detection

    Maoxun Yuan, Yinyan Wang, and Xingxing Wei. Translation, scale and rotation: cross- modal alignment meets rgb-infrared vehicle detection. InEuropean Conference on Computer Vision, pages 509–525. Springer, 2022

    2022

  27. [35]

    Cutmix: Regularization strategy to train strong classifiers with local- izable features

    Sangdoo Yun, Dongyoon Han, Seong Joon Oh, Sanghyuk Chun, Junsuk Choe, and Youngjoon Yoo. Cutmix: Regularization strategy to train strong classifiers with local- izable features. InProceedings of the IEEE/CVF international conference on computer vision, pages 6023–6032, 2019

    2019

  28. [36]

    Guided attentive feature fusion for multispectral pedestrian detection

    Heng Zhang, Elisa Fromont, Sébastien Lefèvre, and Bruno Avignon. Guided attentive feature fusion for multispectral pedestrian detection. InProceedings of the IEEE/CVF winter conference on applications of computer vision, pages 72–80, 2021

    2021

  29. [37]

    Weakly aligned cross-modal learning for multispectral pedestrian detection

    Lu Zhang, Xiangyu Zhu, Xiangyu Chen, Xu Yang, Zhen Lei, and Zhiyong Liu. Weakly aligned cross-modal learning for multispectral pedestrian detection. InProceedings of the IEEE/CVF international conference on computer vision, pages 5127–5137, 2019

    2019

  30. [38]

    Cddfuse: Correlation-driven dual-branch feature decom- position for multi-modality image fusion

    Zixiang Zhao, Haowen Bai, Jiangshe Zhang, Yulun Zhang, Shuang Xu, Zudi Lin, Radu Timofte, and Luc Van Gool. Cddfuse: Correlation-driven dual-branch feature decom- position for multi-modality image fusion. InProceedings of the IEEE/CVF conference on computer vision and pattern ...

    2023

  31. [6857]

    URLhttps://doi.org/10.1145/3545572

    doi: 10.1145/3545572. URLhttps://doi.org/10.1145/3545572

    doi

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.