Pith. sign in

REVIEW 3 major objections 6 minor 46 references

Transformer-Based Dual-Optical Attention Fusion Crowd Head Point Counting and Localization Network

T0 review · 3 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read This paper claims that fusing infrared with visible light, using two attention-style fusion modules plus a spatial-offset augmentation, lowers crowd counting error on UAV RGB-TIR benchmarks, with MAE falling from 10.32 to 7.32 on DroneRGBT.

desk verdict Table 2's headline R-T results don't match any single model in the paper's own ablations, so the central empirical claim is unsubstantiated. read the letter →

arxiv 2505.06937 v1 pith:3RYBTOMV submitted 2025-05-11 cs.CV

Fei Zhou , Yi Li , Mingqing Zhu This is my paper · ORCID
classification cs.CV
keywords crowdcountingRGB-thermalfusionpoint-basedlocalizationattentionUAVimagerymisalignmentaugmentationmulti-kernelMMDtransformer
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

This paper tries to establish that RGB-thermal image fusion, organized by two attention-style fusion modules and trained with a spatial-offset augmentation, makes UAV-view crowd counting and head localization more accurate than single-modality counting, especially at night or in dense scenes. The proposed network, TAPNet, reports mean absolute error 7.32 and F1 0.657 on DroneRGBT and 7.87 and 0.526 on GAIIC2 when both modalities are used, compared with larger errors for RGB-only or TIR-only versions. The claim matters because visible-light-only counters degrade under low light, and RGB-TIR cameras on drones are systematically misaligned; the paper's wager is that both problems can be handled at the feature-fusion level rather than by explicit image registration.

What carries the argument

The carrying mechanism is a pair of fusion modules plus a training-time alignment simulation. DAFP fuses the third and fourth backbon e feature maps of RGB and TIR by concatenating channel-pooled and spatially pooled features and generating per-channel and per-pixel weights through shared MLPs and convolutional layers. AFDF instead fuses early, running both modalities through a Transformer shared layer, a global-structure encoder, a detail encoder, and a hybrid Gaussian-Laplacian multi-kernel MMD term that pulls the modality distributions together in a reproducing kernel Hilbert space. Spatial random offset augmentation shifts the TIR image by random horizontal and vertical offsets in the range [-10, 10] during training, so the network sees misalignment similar to what the validation set contains.

What would settle it

Measure the actual per-pair displacement between RGB and TIR in DroneRGBT and GAIIC2, for example by matching annotated head points or estimating homographies: if typical misalignment exceeds 10 pixels or varies across the image, the augmentation's assumed distribution is wrong, and the reported gains should not transfer to other misaligned datasets. Retraining with offsets drawn from a wider or adaptive distribution and comparing MAE on an independent misaligned set would settle the claim.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that a point-based head detection framework can be made multimodal by adding two complementary fusion paths: DAFP, which applies channel and spatial attention to the last two feature levels of RGB and TIR, and AFDF, an early-fusion encoder-decoder with a Transformer shared layer, global-structure and detail encoders, and hybrid multi-kernel MMD alignment. The paper reports that this combination lowers MAE on DroneRGBT from 10.32 for RGB-only and 13.15 for TIR-only to 7.32 for RGB-T input, and on GAIIC2 from 8.54 and 13.91 to 7.87, with corresponding F1 increases. It also reports that spatial random offset augmentation, shifting the TIR image by up to 10 pixels during training, reduces MAE by about 1.27 and 1.11 for the two fusion modules.

Load-bearing premise

The spatial random offset augmentation assumes that the systematic misalignment between RGB and TIR pairs is a global translation of at most 10 pixels, and that jittering the thermal image by that amount makes training images match the validation distribution; the paper does not measure the actual misalignment or test other shift magnitudes.

Editorial extensions

If this is right

  • With both modalities available, TAPNet reduces MAE on DroneRGBT from 10.32 to 7.32 and on GAIIC2 from 8.54 to 7.87, while raising F1, which makes RGB-TIR counting preferable whenever a thermal camera is available.
  • The auxiliary-point matching framework works best with only positive auxiliary points on fused bi-optical data, suggesting that fused features make negative-point guidance less necessary in this regime.
  • Head-point counting outperforms box counting by 1.66 MAE and 0.032 F1 on DroneRGBT, so point supervision is the more effective label type for dense UAV crowds.
  • Spatial random offset augmentation is a cheap route to misalignment robustness, improving MAE by 1.27 for DAFP and 1.11 for AFDF without adding fusion machinery.
  • Early fusion via AFDF and late attention fusion via DAFP trade off: DAFP is faster and lighter, while AFDF gives higher F1 on the DroneRGBT evaluation, so the choice depends on deployment constraints.

Reading between the lines

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

  • If the actual RGB-TIR misalignment in drone data is well modelled by a global translation of up to 10 pixels, then the augmentation is a near-free substitute for explicit registration; if misalignment is locally varying, AFDF's learned alignment would need to carry the full burden.
  • A natural testable extension is to apply the same hybrid-kernel MMD alignment to other paired modalities, such as RGB-depth or multi-view images, where systematic small shifts between sensors also occur.
  • Because GAIIC2's validation set was manually annotated by the authors, an independent re-annotation or a third-party benchmark would clarify how much of the reported gain is model behaviour versus annotation and matching protocol.
  • The reported F1 values are computed at threshold 0.8; evaluating across thresholds would show whether the localization gain is concentrated in high-confidence heads or uniform over the precision-recall curve.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes TAPNet, a transformer-based RGB-T crowd counting and localization network that combines a dual-optical attention fusion module (DAFP), an adaptive feature decomposition fusion module (AFDF), and a spatial random offset data augmentation strategy. The authors claim state-of-the-art results on DroneRGBT and GAIIC2, with best RGB-T results of 7.32/11.54/0.657 and 7.87/13.25/0.526 in MAE/MSE/F1, respectively. The manuscript includes architecture details, loss formulations, and ablation studies on fusion modules, auxiliary point settings, and the offset augmentation.

Significance. If the reported results were reliable, the paper would make a useful contribution to multimodal crowd counting, particularly for low-light and misaligned RGB-TIR image pairs, and the public release of code would aid reproducibility. The topic is relevant, and the point-based localization with dual-modal fusion is a reasonable direction. However, the central empirical claim is undermined by internal inconsistencies across the tables and by a potentially circular evaluation protocol; the evidence as presented does not establish the stated superiority of the full TAPNet architecture.

major comments (3)
  1. [Section 6, Tables 2, 3, 5, 8] The final R-T row in Table 2 is not traceable to a single well-defined model configuration. On DroneRGBT, Table 2 reports MAE/MSE/F1 = 7.32/11.54/0.657, but Table 3 reports R-T+DAFP = 7.32/11.71/0.697; the MSE and F1 in Table 2 exactly match the (kp,kn) = (1,0) row of Table 5, not the DAFP-only row. On GAIIC2, Table 2's 7.87/13.25/0.526 is identical to the 'AFDF+Spatial Shift' row of Table 8, not to any configuration combining DAFP and AFDF. Thus the headline superiority claim cannot be attributed to the complete TAPNet architecture, and no ablation in the paper reports the full model.
  2. [Sections 4.4 and 5.3] The spatial random offset augmentation is explicitly stated in Section 5.3 to 'enable the validation set to have the same image offset distribution as the training set.' Since the final metrics are reported on that same validation set, the improvement attributed to the augmentation may reflect fitting the validation distribution rather than a generalizable alignment mechanism. The paper provides no measurement of the actual RGB-TIR misalignment in either dataset and no evaluation on a held-out test set. This circularity is a load-bearing issue for the claimed generalization benefit.
  3. [Section 5.3 and Tables 5–6] Hyperparameters for auxiliary point counts (kp, kn) and random ranges (npos, nneg) are selected on the same validation set that is used for the final reported metrics, with no separate validation split. Combined with the internal inconsistency in Table 2, this makes it difficult to rule out overfitting to the benchmark as the source of the reported gains. A separate validation set or cross-validation should be used for hyperparameter selection.
minor comments (6)
  1. [Equation (12)] The equation lists two identical expressions for w'_{c2}; the first should presumably be w'_{c1}.
  2. [Section 5.3] The text refers to 'the spatially randomized offset data enhancement strategy proposed in 2.4'; this should reference Section 4.4.
  3. [Section 4.3] There is a typo: 'Transorfmer shared layer' should read 'Transformer shared layer'.
  4. [Equation (25)] The decomposition loss Lcc uses parameters α and β that are not defined in the surrounding text; please clarify whether these are the same as the modality fusion weights in Equation (13).
  5. [Table 3 caption] The caption says 'Dual Fusion Module' but the table actually compares two separate modules, DAFP and AFDF, evaluated independently; the caption should be clarified.
  6. [References] Several references list only 'et al.' without author names (e.g., [1], [3], [4], [5], [7], [9], [10], [12], [14], [15], [16], [17], [18], [19], [20], [23], [24], [25], [27], [29], [30], [31], [32], [33], [34], [35], [37], [38], [39], [40], [41], [43], [44], [45]); please complete the author lists.

Circularity Check

3 steps flagged · score 6.0 of 10

Headline results are selected from validation-set hyperparameter and ablation searches: the reported DroneRGBT/GAIIC2 numbers coincide with the best auxiliary-point row and an AFDF+Spatial-Shift ablation row, so the claimed TAPNet superiority is partly an artifact of benchmark fitting.

  1. fitted input called prediction [Section 6.2 (Table 5) vs Section 6.1 (Table 2)]
    "We verify the effect of the number of auxiliary positive and negative points on the performance of bi-optical crowd counting in Table 5, and the results show that using only auxiliary positive points for bi-optical data is more likely to utilize the model to select the optimal proposal. ... Table 5: (kp, kn) (1,0) MAE↓ 7.32 MSE↓ 11.54 F1↑ 0.657 ... Table 2: TAPNet (ours) R-T 7.32 11.54 0.657"

    The final DroneRGBT R-T numbers in Table 2 are exactly the best row of Table 5's hyperparameter search over auxiliary point counts, (kp,kn)=(1,0). The configuration is chosen because it is optimal on the validation set, and the same validation-set metrics are then reported as the headline TAPNet result. The 'prediction' is therefore the maximum of a search over the evaluation set; it is not an independent test of a fixed model, and the reported gain over ablations is forced by the selection criterion.

  2. other [Section 6.2 (Table 8) vs Section 6.1 (Table 2, GAIIC2)]
    "Table 8: Ablation Results of Spatial Random Shift Data Augmentation Strategy ... AFDF+Spatial Shift 7.87 13.25 0.526 ... only Table 8 performs the ablation experiments on the GAIIC dataset ... Table 2: TAPNet (ours) ... R-T 7.87 13.25 0.526"

    On GAIIC2, the headline TAPNet R-T row is numerically identical to the AFDF+Spatial Shift ablation row, not to any result for the full DAFP+AFDF architecture described in the abstract. No table reports the complete dual-module model on GAIIC2, so the central empirical claim reduces to a single ablation configuration that is relabeled as the full model; the contribution of the second proposed module is not separately demonstrated.

1 more flagged steps
  1. fitted input called prediction [Sections 4.4 and 5.3]
    "For the offset GAIIC2 dataset, we use the spatially randomized offset data enhancement strategy proposed in 2.4, which enables the validation set to have the same image offset distribution as the training set."

    The spatial random offset augmentation is introduced and its range (Rand(−10,10)) is set so that the training offset distribution matches the validation set's offset distribution, and the final metrics are reported on that same validation set. No independent measurement of real misalignment or a hold-out set with different offsets is provided, so the reported improvement from this strategy is partly a fit to the evaluation distribution rather than a demonstrated generalization to the misalignment problem.

full rationale

The paper is an empirical systems paper, not a formal derivation; there is no self-citation chain or imported uniqueness theorem. However, the central claim 'TAPNet outperforms existing techniques' is not supported by an independent evaluation. The DroneRGBT headline numbers coincide exactly with the best configuration found by the auxiliary-point hyperparameter search on the same validation split, and the GAIIC2 headline numbers coincide exactly with an AFDF+Spatial Shift ablation row; no table reports the complete DAFP+AFDF configuration on a held-out test set. The spatial random offset strategy is likewise justified by matching the validation offset distribution, closing the loop between data-engineering choice and benchmark. These are cases where the reported 'prediction' is by construction the selected row of a validation-set search, so the empirical superiority claim is partially circular. No further circularity, such as self-definitional equations or load-bearing self-citations, was found.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The central claim depends on numerous hand-tuned hyperparameters, the unverified translational-misalignment model underlying the augmentation, and the author-generated GAIIC2 ground truth. No new physical entities are introduced.

free parameters (7)
  • Modality fusion weights α and β = not reported; initialized equal and learned
    Eq. 13 balances RGB and TIR contributions in the DAFP fusion output; learned during training on the evaluation datasets.
  • Point offset scaling γ = 100
    Eq. 1 scales regressed coordinate offsets; set by hand to match dataset coordinate ranges.
  • Hungarian matching weight τ = 2e-2
    Eq. 4 balances Euclidean distance against confidence in the matching cost matrix; tuned on validation.
  • Spatial shift range (Δx, Δy) = uniform [-10, 10] pixels
    Section 4.4; chosen so the training offset distribution matches the validation set, so its benefit is benchmark-specific.
  • Auxiliary point count (k_pos, k_neg) and range (n_pos, n_neg) = (1, 0) and (1, 4)
    Section 6.2; selected by ablation on DroneRGBT and presented as the best configuration.
  • Loss weights β1-β4, γ1-γ2, λ1-λ4 = 2.0, 2.0, 0.1, 1.0; 10, 2; 0.5, 2e-4, 2e-4, 0.2
    Section 5.3; hand-set or tuned on validation; no sensitivity analysis is reported for most of these weights.
  • Number of reference points K = 4
    Section 5.3; chosen based on dataset statistics to ensure the number of proposals exceeds the number of true points.
assumptions (5)
  • domain assumption RGB and TIR image pairs are approximately translationally misaligned with offsets bounded by about 10 pixels.
    Used in Sections 4.4 and 5.3 to justify the spatial random offset augmentation; no measurement of actual misalignment is provided.
  • domain assumption The author-created GAIIC2 validation annotations are correct and comparable to the DroneRGBT labeling protocol.
    Section 5.1 states the GAIIC2 validation set was manually annotated by the authors; all GAIIC2 metrics depend on this ground truth.
  • domain assumption MK-MMD with a convex combination of Gaussian and Laplacian kernels can align cross-modal feature distributions sufficiently for head-point matching.
    Section 4.3 introduces the MK-MMD loss based on DAF-Net; the paper provides no analysis showing that distribution matching corrects geometric misalignment.
  • domain assumption ImageNet pretrained ResNet50 features are suitable for UAV RGB-T head detection.
    Section 4.1 chooses ResNet50 as the backbone based on prior detection works; this transfer assumption is not validated in the UAV RGB-T setting.
  • standard math Hungarian matching on the cost matrix in Eq. 4 yields the correct one-to-one point assignment.
    Standard assignment algorithm used in Section 3.3; not itself a source of novelty.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Transformer-Based Dual-Optical Attention Fusion Crowd Head Point Counting and Localization Network." pith.science (2026). https://pith.science/paper/3RYBTOMV

@misc{pith2026250506937,
  author       = {Pith},
  title        = {Pith review of: Transformer-Based Dual-Optical Attention Fusion Crowd Head Point Counting and Localization Network},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3RYBTOMV}},
  note         = {Machine review of arXiv:2505.06937}
}
read the original abstract

In this paper, the dual-optical attention fusion crowd head point counting model (TAPNet) is proposed to address the problem of the difficulty of accurate counting in complex scenes such as crowd dense occlusion and low light in crowd counting tasks under UAV view. The model designs a dual-optical attention fusion module (DAFP) by introducing complementary information from infrared images to improve the accuracy and robustness of all-day crowd counting. In order to fully utilize different modal information and solve the problem of inaccurate localization caused by systematic misalignment between image pairs, this paper also proposes an adaptive two-optical feature decomposition fusion module (AFDF). In addition, we optimize the training strategy to improve the model robustness through spatial random offset data augmentation. Experiments on two challenging public datasets, DroneRGBT and GAIIC2, show that the proposed method outperforms existing techniques in terms of performance, especially in challenging dense low-light scenes. Code is available at https://github.com/zz-zik/TAPNet

Figures

Figures reproduced from arXiv: 2505.06937 by the authors.

Figure 1
Figure 1. Examples of infrared and visible images. (a) The two rows of people on the left are almost invisible in the visible spectrum under low light conditions, illus￾trating the fact that IR images are more advantageous in low light conditions. (b) Example of RGB-TIR modal mis￾alignment, showing that the modal misalignment problem is more prominent in target detection from the UAV view￾point, where the yellow and red boxes… view at source ↗
Figure 2
Figure 2. Overall architecture of TAPNet. We first extract the image feature representation {FR1, ..., FR4} and {FT1, ..., FT4} separately using the ResNet50 backbone. Then, a bi-optical attention fusion module is applied to the last two layers of features to fuse the features. Subsequently, the fused two layers of features {F3, F4} are passed through an adaptive spatial pyramid pooling (ASPP) module and implicit feature inte… view at source ↗
Figure 3
Figure 3. Adaptive Fusion Architecture Diagram. The module consists of an encoder and decoder, respectively, and a domain-adaptive layer structure based on hybrid kernel functions. The difference with [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 37 canonical work pages

  1. [1]

    Frequency bias in neural networks for input of non-uniform density

    Ronen Basri and et al. Frequency bias in neural networks for input of non-uniform density. In International Conference on Machine Learning, 2020. 12

    work page 2020

  2. [2]

    Improving point-based crowd counting and localization based on auxiliary point guidance

    I-Hsiang Chen and et al. Improving point-based crowd counting and localization based on auxiliary point guidance. arXiv, page abs/2405.10589, 2024

    arXiv 2024

  3. [3]

    The effectiveness of a simplified model structure for crowd counting

    Lei Chen and et al. The effectiveness of a simplified model structure for crowd counting. arXiv, 2024

    work page 2024

  4. [4]

    Simple Training Strategies and Model Scaling for Object Detection

    Xianzhi Du and et al. Simple training strategies and model scaling for object detection. In arXiv, page abs/2107.00057, 2021

    work page Pith review arXiv 2021

  5. [5]

    Cross-modality fusion trans- former for multispectral object detection

    Qing Fu Fang and et al. Cross-modality fusion trans- former for multispectral object detection. In arXiv, page abs/2111.00273, 2021

    arXiv 2021

  6. [6]

    Lraf-net: Long-range attention fusion network for visible–infrared object detection.IEEE Transac- tions on Neural Networks and Learning Systems, 35:13232– 13243, 2023

    Haolong Fu and et al. Lraf-net: Long-range attention fusion network for visible–infrared object detection.IEEE Transac- tions on Neural Networks and Learning Systems, 35:13232– 13243, 2023

    work page 2023

  7. [7]

    Learning independent instance maps for crowd localization

    Junyu Gao and et al. Learning independent instance maps for crowd localization. In arXiv, page abs/2012.04164, 2020

    arXiv 2012

  8. [8]

    Simple copy-paste is a strong data augmentation method for instance segmentation

    Golnaz Ghiasi and et al. Simple copy-paste is a strong data augmentation method for instance segmentation. In CVPR, pages 2917–2927, 2020

    work page 2020

Show all 46 references
  1. [9]

    Dpdetr: Decoupled position detection transformer for infrared-visible object detection

    Junjie Guo and et al. Dpdetr: Decoupled position detection transformer for infrared-visible object detection. In arXiv, page abs/2408.06123, 2024

    2024

  2. [10]

    Composition loss for counting, den- sity map estimation and localization in dense crowds

    Haroon Idrees and et al. Composition loss for counting, den- sity map estimation and localization in dense crowds. In arXiv, page abs/1808.01050, 2018

    2018 arXiv

  3. [11]

    Attention scaling for crowd counting

    Xiaoheng Jiang, Li Zhang, Mingliang Xu, Tianzhu Zhang, Pei Lv, Bing Zhou, Xin Yang, and Yanwei Pang. Attention scaling for crowd counting. In CVPR, 2020

    2020

  4. [12]

    Multimodal crowd counting with pix2pix gans

    Muhammad Asif Khan and et al. Multimodal crowd counting with pix2pix gans. In VISIGRAPP : VISAPP, 2024

    2024

  5. [13]

    The hungarian method for the assignment problem

    Harold W Kuhn. The hungarian method for the assignment problem. Naval Research Logistics (NRL), 52, 1955

    1955

  6. [14]

    Illumination-aware faster r-cnn for robust multispectral pedestrian detection

    Chengyang Li and et al. Illumination-aware faster r-cnn for robust multispectral pedestrian detection. Pattern Recognit., 85:161–171, 2018

    2018

  7. [15]

    Confidence-aware fusion using dempster- shafer theory for multispectral pedestrian detection

    Qing Li and et al. Confidence-aware fusion using dempster- shafer theory for multispectral pedestrian detection. IEEE TMM, 25:3420–3431, 2023

    2023

  8. [16]

    Stabilizing multispectral pedestrian de- tection with evidential hybrid fusion

    Qing Li and et al. Stabilizing multispectral pedestrian de- tection with evidential hybrid fusion. IEEE Transactions on Circuits and Systems for Video Technology , 34:3017–3029, 2024

    2024

  9. [17]

    Focal inverse distance transform maps for crowd localization

    Dingkang Liang and et al. Focal inverse distance transform maps for crowd localization. IEEE TMM , 25:6040–6052, 2021

    2021

  10. [18]

    An end-to-end transformer model for crowd localization

    Dingkang Liang and et al. An end-to-end transformer model for crowd localization. InarXiv, page abs/2202.13065, 2022

    2022 arXiv

  11. [19]

    Lin and A

    W. Lin and A. B. Chan. Optimal transport minimization: Crowd localization on density maps for semi-supervised counting. In CVPR, pages 21663–21673, 2023

    2023

  12. [20]

    Point-query quadtree for crowd counting, localization, and more

    Chengxin Liu and et al. Point-query quadtree for crowd counting, localization, and more. In ICCV, pages 1676– 1685, 2023

    2023

  13. [21]

    Multispectral deep neural networks for pedes- trian detection

    Jingjing Liu. Multispectral deep neural networks for pedes- trian detection. 2016

    2016

  14. [22]

    Leveraging self-supervision for cross-domain crowd counting

    Weizhe Liu, Nikita Durasov, and Pascal Fua. Leveraging self-supervision for cross-domain crowd counting. InCVPR, 2022

    2022

  15. [23]

    Point in, box out: Beyond counting persons in crowds

    Yuting Liu and et al. Point in, box out: Beyond counting persons in crowds. In CVPR, pages 6462–6471, 2019

    2019

  16. [24]

    Implicit neural representation in med- ical imaging: A comparative survey

    Amir Molaei and et al. Implicit neural representation in med- ical imaging: A comparative survey. In IEEE/CVF Interna- tional Conference on Computer Vision Workshops (ICCVW), pages 2373–2383, 2023

    2023

  17. [25]

    Deepsdf: Learning continuous signed distance functions for shape representation

    Jeong Joon Park and et al. Deepsdf: Learning continuous signed distance functions for shape representation. InCVPR, pages 165–174, 2019

    2019

  18. [26]

    Rgb-t crowd counting from drone: A benchmark and mmccn network

    Tao Peng, Qing Li, and Pengfei Zhu. Rgb-t crowd counting from drone: A benchmark and mmccn network. InComputer Vision – ACCV 2020: 15th Asian Conference on Computer Vision, Kyoto, Japan, November 30 – December 4, 2020, Re- vised Selected Papers, Part VI , page 497–513, Berlin,...

    2020

  19. [27]

    On the spectral bias of neural net- works

    Nasim Rahaman and et al. On the spectral bias of neural net- works. In International Conference on Machine Learning , 2018

    2018

  20. [28]

    Faster r-cnn: Towards real-time ob- ject detection with region proposal networks

    Shaoqing Ren and et al. Faster r-cnn: Towards real-time ob- ject detection with region proposal networks. IEEE TPAMI, 39:1137–1149, 2015

    2015

  21. [29]

    Locate, size, and count: Accu- rately resolving people in dense crowds via detection

    Deepak Babu Sam and et al. Locate, size, and count: Accu- rately resolving people in dense crowds via detection. IEEE TPAMI, 43:2739–2751, 2019

    2019

  22. [30]

    Icafusion: Iterative cross-attention guided feature fusion for multispectral object detection

    Jifeng Shen and et al. Icafusion: Iterative cross-attention guided feature fusion for multispectral object detection. In arXiv, page abs/2308.07504, 2023

    2023 arXiv

  23. [31]

    Rethinking counting and localization in crowds: A purely point-based framework

    Qingyu Song and et al. Rethinking counting and localization in crowds: A purely point-based framework. In ICCV, pages 3345–3354, 2021

    2021

  24. [32]

    Fourier features let networks learn high frequency functions in low dimensional domains

    Matthew Tancik and et al. Fourier features let networks learn high frequency functions in low dimensional domains. In arXiv, page abs/2006.10739, 2020

    2006 arXiv

  25. [33]

    Counting in dense crowds using deep features

    Karunya Tota and Haroon Idrees. Counting in dense crowds using deep features. 2015

    2015

  26. [34]

    Multispectral pedestrian detection using deep fusion convolutional neural networks

    J ¨org Wagner and et al. Multispectral pedestrian detection using deep fusion convolutional neural networks. In The Eu- ropean Symposium on Artificial Neural Networks, 2016

    2016

  27. [35]

    A generalized loss function for crowd counting and localization

    Jia Wan, Ziquan Liu, and Antoni B Chan. A generalized loss function for crowd counting and localization. In CVPR, 2021

    2021

  28. [36]

    Dronenet: Rescue drone-view object detection

    Xiandong Wang and et al. Dronenet: Rescue drone-view object detection. Drones, 2023

    2023

  29. [37]

    Daf-net: A dual-branch feature decom- position fusion network with domain adaptive for infrared and visible image fusion

    Jian Xu and Xin He. Daf-net: A dual-branch feature decom- position fusion network with domain adaptive for infrared and visible image fusion. In arXiv, page abs/2409.11642, 2024

    2024 arXiv

  30. [38]

    Ultrasr: Spatial encoding is a missing key for implicit image function-based arbitrary-scale super- resolution

    Xingqian Xu and et al. Ultrasr: Spatial encoding is a missing key for implicit image function-based arbitrary-scale super- resolution. In arXiv, page abs/2103.12716, 2021

    2021 arXiv

  31. [39]

    Baanet: Learning bi-directional adaptive attention gates for multispectral pedestrian detec- tion

    Xiaoxiao Yang and et al. Baanet: Learning bi-directional adaptive attention gates for multispectral pedestrian detec- tion. In ICRA, pages 2920–2926, 2021. 13

    2021

  32. [40]

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

    Maoxun Yuan and et al. Translation, scale and rotation: Cross-modal alignment meets rgb-infrared vehicle detection. In arXiv, page abs/2209.13801, 2022

    2022 arXiv

  33. [41]

    Improving rgb-infrared object detec- tion with cascade alignment-guided transformer.Inf

    Maoxun Yuan and et al. Improving rgb-infrared object detec- tion with cascade alignment-guided transformer.Inf. Fusion, 105:102246, 2024

    2024

  34. [42]

    C²former: Calibrated and complementary transformer for rgb-infrared object de- tection

    Maoxun Yuan and Xingxing Wei. C²former: Calibrated and complementary transformer for rgb-infrared object de- tection. IEEE Transactions on Geoscience and Remote Sens- ing, 62:1–12, 2023

    2023

  35. [43]

    Weakly aligned feature fusion for mul- timodal object detection

    Lu Zhang and et al. Weakly aligned feature fusion for mul- timodal object detection. IEEE Transactions on Neural Net- works and Learning Systems, PP, 2021

    2021

  36. [44]

    Single-image crowd counting via multi-column convolutional neural network

    Yingying Zhang and et al. Single-image crowd counting via multi-column convolutional neural network. InCVPR, pages 589–597, 2016

    2016

  37. [45]

    Improving multispectral pedestrian detection by addressing modality imbalance problems

    Kailai Zhou and et al. Improving multispectral pedestrian detection by addressing modality imbalance problems. In ECCV, 2020

    2020

  38. [46]

    Y . Zhu, X. Sun, M. Wang, and H. Huang. Multi-modal feature pyramid transformer for rgb-infrared object detec- tion. IEEE Transactions on Intelligent Transportation Sys- tems, 24(9):9984–9995, 2023. 14

    2023

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.