Pith. sign in

REVIEW 3 major objections 5 minor 32 references

FressDet, a C4-rotation-equivariant detector that treats spectral bands as a continuous ordered axis, matches or beats much larger multispectral detectors on five benchmarks while using 93% fewer parameters.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 08:27 UTC pith:5I33ZNUZ

load-bearing objection A useful, well-engineered multispectral detector with honest supplementary proofs, but the 'fully equivariant' label runs ahead of the rectangular input sizes used in most benchmarks. the 3 major comments →

arxiv 2607.05148 v2 pith:5I33ZNUZ submitted 2026-07-06 cs.CV

Fully Rotation-Equivariant Spectral-Spatial Learning for Multispectral Object Detection

Peng Zhang , Tingfa Xu , Shuaihao Han , Jianan Li This is my paper
classification cs.CV
keywords multispectral object detectionrotation equivariancegroup equivariant networksimplicit neural representationscontinuous spectral modelingspectral-spatial fusionoriented object detectionC4 symmetry
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper sets out to show that a multispectral object detector does not need to be large if it respects two structures in the data: the 90-degree rotational symmetry of the image plane and the ordered continuity of the spectral dimension. It introduces FressDet, a single-stream detector whose features carry an explicit orientation index (the C4 group) and whose spectral channels are resampled through a learned monotone warp, so that rotating the input merely permutes the output features. Across five public benchmarks—aerial, hyperspectral, and RGB-thermal—the authors report state-of-the-art accuracy with a 2.3M-parameter model, about 93% fewer parameters than the previous best method, and the smallest drop under 1-degree input rotations. The claim is that explicit symmetry and continuous spectral modeling, rather than scale, are what buy accuracy and stability in this setting.

Core claim

The central discovery claimed is that full C4 rotation equivariance can be combined with continuous, order-preserving spectral resampling in a single-stream detector without parameter replication, and that this combination yields state-of-the-art accuracy at a fraction of the parameter count of prior work. SpeIW converts the discrete band axis into a coordinate-conditioned implicit field whose positive increments are cumulatively summed, guaranteeing that spectral channel order is preserved under resampling. ReCoW adaptively gates spectral and spatial cues by their rotation-invariant agreement, and the oriented-aware head reads out class, box, and angle predictions from group-indexed feature

What carries the argument

The load-bearing object is the C4 regular representation carried by every feature map: a 90-degree input rotation becomes a cyclic permutation of the orientation axis plus a spatial rotation, and each module is built to commute with that action. SpeIW builds a continuous spectral warp: Fourier-encoded spectral coordinates feed a shared MLP that outputs basis vectors, inner products with a spatial latent code give per-pixel offsets, and softplus-plus-cumulative-sum forces strictly increasing resampling indices. ReCoW routes two prototypes (soft spectral, hard spatial) and weights a residual update by their group-averaged cosine agreement; the oriented head uses group-pooled classification and

Load-bearing premise

The proof that every component is C4-equivariant requires feature lattices and ReCoW window partitions that map to themselves under 90-degree rotation, which holds for square grids; several reported benchmarks use rectangular input sizes and the paper does not describe a square-cropping or padding step, so if that condition is not met in the implementation, the 'fully rotation-equivariant' guarantee does not apply to those experiments.

What would settle it

Take a MODA image at its native 1216x928 resolution, rotate it by exactly 90 degrees, and measure the normalized equivariance error (Eq. 2 in the supplement) at each backbone stage under the released code. If the error is orders of magnitude above floating-point round-off, or if rotating the input by 90 degrees does not produce a correspondingly rotated output feature map, the full-equivariance claim fails for the reported setting; the same check with square-padded inputs would isolate whether the lattice condition is the culprit.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • SpeIW's shared implicit field can be queried at any spectral coordinate, so the same warp machinery applies to inputs with different or variable numbers of bands without scaling parameter count with band count.
  • On MODA, FressDet raises mAP50 from 69.0 (OSSDet) to 73.1 while shrinking parameters from 36.6M to 2.3M and FLOPs from 263G to 31.7G; the largest gains are on small classes (bike +16.7, pedestrian +26.8 mAP50).
  • Replacing the spectral operator with four deformable-convolution variants drops mAP by 1.1 to 1.5 points, supporting the claim that the order-preserving continuous warp, not equivariance alone, carries part of the gain.
  • Trained from scratch without rotation augmentation, FressDet keeps lower equivariance error than a non-equivariant baseline at every backbone stage and degrades less than competitors under 1-degree-step rotations of MODA.
  • The SpeIWNet backbone transfers to five existing detectors, each gaining accuracy while shrinking parameter count, so the contribution is not tied to one head design.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Because the spectral warp is coordinate-conditioned and the coefficients are pixel-conditioned, the same field could be reused at inference on images with different spectral calibration or band order; a test would be zero-shot evaluation on a dataset with a different band set.
  • The rectangular-input gap between the square-lattice proof and the MODA, HOD3K, and LLVIP resolutions suggests a concrete fix—symmetric zero-padding to a square before the lifting layer—and a measurable prediction: after that fix, exact 90-degree equivariance error on those datasets should drop to floating-point level.
  • The 93% parameter reduction implies that much of the capacity in conventional detectors is spent memorizing rotated variants of features; if so, similar compression ratios should appear when this architecture is applied to other rotation-heavy dense tasks, such as aerial segmentation.
  • ReCoW's soft/hard routing asymmetry is a design principle that could generalize beyond multispectral fusion: use soft assignment where cues mix continuously, hard assignment where boundaries matter.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper introduces FressDet, a C4-equivariant single-stream multispectral object detector. Its three components are: SpeIW, an implicit spectral warp that resamples the channel/spectral axis with a monotone order-preserving map; ReCoW, a reliability-aware fusion module that combines spectral and spatial routing through an agreement-weighted residual gate; and an oriented-aware head that uses group-indexed features to predict oriented boxes. Experiments on MODA, HOD3K, DroneVehicle and, in the supplement, LLVIP and VEDAI report state-of-the-art mAP with a 2.3M-parameter model. The supplement provides conditional equivariance proofs with explicit caveats about numerical precision, lattice conditions, unique routing, and angle-branch degeneracy.

Significance. If the claims hold, the result is significant: it unifies continuous spectral modeling with group-equivariant detection and shows that a compact equivariant model with 2.3M parameters can match or exceed much larger baselines on several multispectral benchmarks. The manuscript's strengths are its unusually candid supplement — acknowledging that full equivariance is algebraic and conditional on stated compatibility assumptions — the released code, the controlled RE-DCN spectral-operator ablation, and the cross-detector backbone transfer experiments. The main unresolved issue is that the stated proof conditions are not met by three of the five reported evaluation settings, so the central 'fully rotation-equivariant' claim is not established for the headline experiments as described.

major comments (3)
  1. [Supp. S1.2/S3.4 vs S5.3] The end-to-end C4-equivariance proof is conditional on feature lattices and ReCoW window partitions being invariant under 90 degree rotations; S1.2 explicitly reasons about square feature grids with symmetric constant padding, and S3.4 lists lattice compatibility as condition (i). However, S5.3 reports rectangular training inputs for MODA (1216x928), HOD3K (1024x512), and LLVIP (1280x1024), and no square-crop or square-pad step is described in the main text or supplement. A 90 degree rotation of a 1216x928 tensor yields a 928x1216 tensor, so the input domain is not closed under the claimed group action. The main 'fully rotation-equivariant' statement therefore does not cover three of the five benchmarks, including the headline MODA result. The paper must either add an explicit, symmetry-compatible square padding/cropping step and prove it, or qualify the guarantee to the square-lattice s
  2. [Abstract and Section 3.1] The abstract and introduction claim equivariance 'under arbitrary in-plane rotations,' but the algebraic framework is defined only for the cyclic group C4. Arbitrary-angle behavior is measured empirically in Fig. 1(b) at 1 degree increments (see also the rotation-robustness paragraph at the end of Supp. S3.4), not proven. The exact guarantee should be stated as C4 equivariance; arbitrary rotations should be described as empirical robustness, not as part of the algebraic equivariance claim.
  3. [Section 4.3, Fig. 7] The equivariance-error measurement is presented as evidence for the central property, but the protocol is underspecified. For rectangular inputs, 'rotating the input before' changes the tensor shape, so the RMS residual between the two computation paths is not defined without additional canvas manipulation. The paper does not state how the measurement handles this, and the figure therefore cannot be used to verify the algebraic guarantee on the actual benchmark geometry. Please specify the exact protocol, including padding or cropping, or restrict the claim to the square-lattice case.
minor comments (5)
  1. [Section 3.4, Eq. (16c)] The angle-branch shift law in Eq. (16c) has a nonzero-resultant condition that is only disclosed in Supp. S3.3. State this degeneracy condition next to Eq. (16c) to avoid overclaiming.
  2. [Supp. S7] The supplement repeatedly refers to 'Tab. 6 of the main paper' for the RE-DCN spectral-operator ablation. The correct table in the main text is Table 5, not Table 6.
  3. [Supp. S3.2] The sentence contrasting the released gate with 'the convex interpolation beta*S + (1-beta)*T stated in the earlier supplement' is confusing and self-referential. The supplement should present a single canonical formula and remove references to earlier versions.
  4. [Fig. 7] Define the equivariance-error metric and normalization in the main text, and state how rotated non-square canvases are compared. Currently the y-axis scale and the input preprocessing are unclear.
  5. [Table 1 caption] The abbreviation 'Awi.' is not expanded in the caption; use 'Awning-truck' or add a footnote.

Circularity Check

0 steps flagged

No circularity: the equivariance theorems and SpeIW monotone-warp construction are self-contained; the same-group MODA/OSSDet comparisons are not load-bearing.

full rationale

The derivation chain is not circular. SpeIW's order-preserving warp is a construction, not a fitted target: Eqs. (8a)-(8b) define positive softplus increments and cumulative-sum normalization, so monotonicity holds by definition, and the paper says so explicitly ('we obtain an order-preserving mapping by construction'). Equivariance of SpeIW is proven compositionally from the regular representation of z(x,g) and the shared coordinate network, with interpolation only along the spectral axis (Supp. S3.1). ReCoW's equivariance is also proven under explicit lattice, uniqueness, and finiteness assumptions; the supplement honestly states the conditions and floating-point caveats (Supp. S1.1, S3.2, S3.3), and the end-to-end statement (Supp. S3.4) is conditional rather than a claim forced by definition. The rectangular-input mismatch noted by the reader is a correctness/experimental-validity concern, not circularity: the proof's assumptions are not defined in terms of the target result, and no equation is silently reused as its own conclusion. The SOTA claim is supported against a wide set of external baselines (YOLO11m, O-DETR, WaveMamba, RTMDet, etc.); OSSDet and MODA are same-group citations, but the core accuracy and parameter-count comparisons are arithmetic and independently verifiable, not a fitted-input-called-prediction reduction. There is no imported uniqueness theorem and no ansatz smuggled via self-citation. Therefore no circular step meets the evidentiary bar.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The central claim rests on standard approximation theory; on the C4-square-lattice condition that is not met by several rectangular benchmark inputs; on constant zero-padding equivariance; and on non-degeneracy conditions (unique routing winner, nonzero circular resultant). These are stated in the paper, but the first is not reconciled with the actual data shapes.

free parameters (5)
  • Rotation group order |G| = C4 = C4 (N=4)
    Ablation Table 6: C8 gives +0.1 mAP50 at 2x FLOPs; all 'fully rotation-equivariant' claims are for discrete C4, not continuous SO(2).
  • Fourier frequency count L in SpeIW = L=8
    Ablation Table 7: chosen on MODA; impacts spectral high-frequency modeling and overfitting.
  • ReCoW window size and prototype grid = W=4, K=16 (4x4)
    Ablation Table 10: chosen on MODA; 5x5 overfits, 2x2 under-resolves.
  • Numerical offsets eps_s, eps_d in monotonic mapping = not reported (positive)
    Eq. (8): introduced to guarantee positive increments and endpoint below C-1; values not specified, but they affect the warp and the implementation-level order guarantee.
  • Learnable routing scales tau and gamma in ReCoW = learned; final values not reported
    Eqs. (10a)-(12a): temperature and inverse bandwidth are learned from data; no values or constraints reported.
axioms (5)
  • standard math Universal approximation theorem: MLPs can approximate continuous functions on compact domains.
    Section 3.2 and Introduction justify the implicit spectral field; standard, but the guarantee is one of existence, not of the trained network's fidelity.
  • domain assumption C4 spatial action maps the feature lattice to itself (square grids, compatible windows).
    Supplementary S1.2 and S3.4 impose this for equivariance proof; contradicted by rectangular benchmark inputs (MODA 1216x928, HOD3K 1024x512, LLVIP 1280x1024) unless an unstated square-padding step is applied.
  • domain assumption Constant zero padding is C4-equivariant under the regular representation.
    Supplementary Eq. (3); true for square padded maps, but relies on symmetric constant extension and fails for non-square maps without explicit alignment.
  • domain assumption Hard spatial routing has a unique winning prototype and the angle branch resultant is nonzero.
    Supplementary S3.2-S3.3: exact ties and zero/near-zero resultant are resolved by implementation order; the proof excludes these cases.
  • standard math Linear interpolation along spectral channels commutes with the group action.
    Supplementary Eq. (12); correct because interpolation only gathers along the channel axis with equivariant indices; requires indices to transform as Eq. (10).

pith-pipeline@v1.3.0-alltime-deepseek · 29692 in / 15609 out tokens · 151627 ms · 2026-08-02T08:27:57.297267+00:00 · methodology

0 comments
read the original abstract

Existing multispectral detectors are limited by discrete spectral processing, a scale-dependent shift in the relative reliability of spectral and spatial cues across pyramid levels, and the lack of explicit rotation-equivariant geometric priors for arbitrarily oriented objects. To tackle these limitations, we propose FressDet, a fully rotation-equivariant spectral-spatial learning framework for multispectral object detection, capable of capturing the continuous, ordered nature of spectral structure and enabling reliable spectral-spatial fusion across pyramid levels under arbitrary in-plane rotations. FressDet integrates three complementary components. Spectral Implicit Warp (SpeIW) enables query-based spectral resampling via a coordinate-conditioned implicit field, yielding a monotone, order-preserving warp. Rotation-Equivariant Consistency Weighting (ReCoW) adaptively fuses spectral and spatial branches based on branch reliability, reinforcing informative cues while suppressing noise across pyramid levels. The oriented-aware head exploits group-indexed features to stably predict oriented objects without parameter replication. Taken together, FressDet learns more discriminative and robust spectral-spatial representations even under rotational perturbations. By achieving state-of-the-art performance with 93% fewer parameters on five public benchmarks, FressDet demonstrates its effectiveness and generalizability.

Figures

Figures reproduced from arXiv: 2607.05148 by Jianan Li, Peng Zhang, Shuaihao Han, Tingfa Xu.

Figure 1
Figure 1. Figure 1: (a) Transformation consistency in vanilla (V-CNN) and equivariant (Eq-CNN) networks. Te and Tr denote the identity and a rotation, f is a feature extractor, and Z0, Z1 are features of I0 =Te(I0) and I1 =Tr(I0). Ideally T −1 r (Z1)≈Z0; the error map confirms Eq-CNN yields lower transformation error. However, existing methods have not yet unified continuous spectral modeling with rotation-equivariant geometr… view at source ↗
Figure 2
Figure 2. Figure 2: The overall framework of FressDet. (a) SpeIW resamples spectral features via a continuous, order-preserving warp. (b) ReCoW adaptively fuses spectral and spatial branches based on branch reliability. (c) The oriented-aware head stably predicts ori￾ented objects without parameter replication. product of translations and orthogonal transformations. By restricting to the cyclic subgroup CN = {rk | k = 0, . . … view at source ↗
Figure 3
Figure 3. Figure 3: Visualization comparison of detection results and feature maps obtained by different methods on MODA. Please zoom in to see details [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Detection results on HOD3K. Red boxes indicate the ground truth, while green boxes denote the predictions [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: (a) Effectiveness visualization of key components. (b) 2D t-SNE [34] visualiza￾tion of feature distributions on MODA [PITH_FULL_IMAGE:figures/full_fig_p013_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: SpeIWNet backbone generaliza￾tion across detectors on MODA. Error mAP (%) 100 10−1 10−2 10−3 10−4 10−6 10−7 10−8 10−5 0.6 0.5 0.4 0.3 0.2 1 3 5 7 9 11 13 15 17 19 Epoch S0 S1 S2 S3 S4 S0 S1 S2 S3 S4 OSSDet FressDet FressDet OSSDet [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗
Figure 1
Figure 1. Figure 1: Detection results on LLVIP. Red boxes indicate the ground truth, while green boxes denote the predictions [PITH_FULL_IMAGE:figures/full_fig_p030_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of spectral resampling strategies. (a) Regular spectral grid with uniform channel spacing. (b) RE-DCN applies discrete offsets to each channel anchor; independent offsets can violate spectral ordering (red cross). (c) SpeIW learns a con￾tinuous monotone warp that preserves spectral order by construction. S7.2 Four RE-DCN Variants All four variants preserve the same backbone topology and differ o… view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of multiscale feature maps. SLBAF-Net C2Former-S2ANet WaveMamba FressDet (Ours) [PITH_FULL_IMAGE:figures/full_fig_p035_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Detection results on DroneVehicle. Red boxes indicate the ground truth, while green boxes denote the predictions. Discussion. The above proof also clarifies the difference between rotation equiv￾ariance and spectral continuity. RE-DCN variants preserve rotation equivariance because the deformation is conditioned by equivariant features and the resam￾pling is confined to the spectral dimension. However, unl… view at source ↗
Figure 5
Figure 5. Figure 5: demonstrates the rotation equivariance property of FressDet by visualizing feature maps under input rotations of 0 ◦ , 90◦ , 180◦ , and 270◦ . As the input image rotates, the corresponding feature maps rotate accordingly, confirming that FressDet maintains strict rotation equivariance throughout the network. (a) Original image (b) Original image rotated by 90° (c) Original image rotated by 180° (c) Origina… view at source ↗

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

32 extracted references

  1. [1]

    In: CVPR (2018)

    Cai, Z., Vasconcelos, N.: Cascade R-CNN: Delving into high quality object detec- tion. In: CVPR (2018)

    2018

  2. [2]

    In: CVPR

    Cao, Y., Bin, J., Hamari, J., Blasch, E., Liu, Z.: Multimodal object detection by channel switching and spatial attention. In: CVPR. pp. 403–411 (2023)

    2023

  3. [3]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition

    Chen, C., Qi, J., Liu, X., Bin, K., Fu, R., Hu, X., Zhong, P.: Weakly misalignment- free adaptive feature alignment for UAVs-based multimodal object detection. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recog- nition. pp. 26836–26845 (2024)

    2024

  4. [4]

    In: ICML (2016)

    Cohen, T.S., Welling, M.: Group equivariant convolutional networks. In: ICML (2016)

    2016

  5. [5]

    In: 2017 IEEE International Conference on Computer Vision (ICCV)

    Dai, J., Qi, H., Xiong, Y., Li, Y., Zhang, G., Hu, H., Wei, Y.: Deformable con- volutional networks. In: 2017 IEEE International Conference on Computer Vision (ICCV). pp. 764–773 (2017).https://doi.org/10.1109/ICCV.2017.89

    doi:10.1109/iccv.2017.89 2017

  6. [6]

    IEEE Transactions on Neural Networks and Learning Systems35(10), 13232–13245 (2024)

    Fu, H., Wang, S., Duan, P., Xiao, C., Dian, R., Li, S., Li, Z.: LRAF-Net: Long-range attention fusion network for visible–infrared object detection. IEEE Transactions on Neural Networks and Learning Systems35(10), 13232–13245 (2024)

    2024

  7. [7]

    IEEE TPAMI46(10), 6577–6593 (2024)

    Fu, J., Xie, Q., Meng, D., Xu, Z.: Rotation equivariant proximal operator for deep unfolding methods in image restoration. IEEE TPAMI46(10), 6577–6593 (2024)

    2024

  8. [8]

    In: CVPR

    Girdhar, R., El-Nouby, A., Liu, Z., Singh, M., Alwala, K.V., Joulin, A., Misra, I.: ImageBind: One embedding space to bind them all. In: CVPR. pp. 15180–15190 (June 2023)

    2023

  9. [9]

    arXiv preprint arXiv:2512.09489 (2025)

    Han, S., Xu, T., Liu, P., Li, J.: MODA: The first challenging benchmark for multi- spectral object detection in aerial images. arXiv preprint arXiv:2512.09489 (2025)

    arXiv 2025

  10. [10]

    In: Proceedings of the 31st ACM International Con- ference on Multimedia

    He, X., Tang, C., Zou, X., Zhang, W.: Multi-spectral object detection via cross- modal conflict-aware learning. In: Proceedings of the 31st ACM International Con- ference on Multimedia. pp. 1465–1474 (2023)

    2023

  11. [11]

    IEEE Transactions on Geoscience and Remote Sensing61, 1–13 (2023).https://doi.org/10.1109/ TGRS.2023.3307288

    He, X., Tang, C., Liu, X., Zhang, W., Sun, K., Xu, J.: Object detection in hyper- spectral image via unified spectral–spatial feature aggregation. IEEE Transactions on Geoscience and Remote Sensing61, 1–13 (2023).https://doi.org/10.1109/ TGRS.2023.3307288

    arXiv 2023

  12. [12]

    In: ICCV

    Jia, X., Zhu, C., Li, M., Tang, W., Zhou, W.: LLVIP: A visible-infrared paired dataset for low-light vision. In: ICCV. pp. 3496–3504 (2021)

    2021

  13. [13]

    Ultralytics (2023), available athttps: //github.com/ultralytics/ultralytics

    Jocher, G., Chaurasia, A., Qiu, J.: YOLOv8. Ultralytics (2023), available athttps: //github.com/ultralytics/ultralytics. Accessed 5 March 2026

    2023

  14. [14]

    In: AAAI

    Lee, C., Son, J., Shon, H., Jeon, Y., Kim, J.: FRED: Towards a full rotation- equivariance in aerial image object detection. In: AAAI. pp. 2883–2891 (2024). https://doi.org/10.1609/aaai.v38i4.28069

    doi:10.1609/aaai.v38i4.28069 2024

  15. [15]

    In: Computer Vision – ECCV 2014

    Lin, T.Y., Maire, M., Belongie, S., Hays, J., Perona, P., Ramanan, D., Dollár, P., Zitnick, C.L.: Microsoft COCO: Common objects in context. In: Computer Vision – ECCV 2014. pp. 740–755 (2014).https://doi.org/10.1007/978-3-319-10602- 1_48

    doi:10.1007/978-3-319-10602- 2014

  16. [16]

    IEEE Geoscience and Remote Sensing Letters21, 1–5 (2024)

    Nie,J.,Sun,H.,Sun,X.,Ni,L.,Gao,L.: Cross-modalfeaturefusionandinteraction strategy for cnn-transformer-based object detection in visual and infrared remote sensing imagery. IEEE Geoscience and Remote Sensing Letters21, 1–5 (2024)

    2024

  17. [17]

    Journal of Visual Communication and Image Representation 34, 187–203 (2016) Supplementary Material 21

    Razakarivony, S., Jurie, F.: Vehicle detection in aerial imagery: A small target detection benchmark. Journal of Visual Communication and Image Representation 34, 187–203 (2016) Supplementary Material 21

    2016

  18. [18]

    Pattern Recognition (2024)

    Shen, J., Chen, Y., Liu, Y., Zuo, X., Fan, H., Yang, W.: ICAFusion: Iterative cross-attention guided feature fusion for multispectral object detection. Pattern Recognition (2024)

    2024

  19. [19]

    IEEE TCSVT32(10), 6700–6713 (2022)

    Sun,Y.,Cao,B.,Zhu,P.,Hu,Q.:Drone-basedRGB-infraredcross-modalityvehicle detection via uncertainty-aware learning. IEEE TCSVT32(10), 6700–6713 (2022). https://doi.org/10.1109/TCSVT.2022.3168279

    arXiv 2022

  20. [20]

    IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing18, 2749–2765 (2025)

    Tian, S., Wang, L., Cao, L., Kang, L., Sun, X., Tian, J., Xing, X., Shen, B., Fan, C., Du, K., Fu, C., Zhang, Y.: A dynamic cascade cross-modal coassisted network for aav image object detection. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing18, 2749–2765 (2025)

    2025

  21. [21]

    In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition

    Wang, W., Dai, J., Chen, Z., Huang, Z., Li, Z., Zhu, X., Hu, X., Lu, T., Lu, L., Li, H., Wang, X., Qiao, Y.: InternImage: Exploring large-scale vision foundation models with deformable convolutions. In: Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition. pp. 14408–14419 (2023). https://doi.org/10.1109/CVPR52729.2023.01385

    arXiv 2023

  22. [22]

    In: NeurIPS (2019)

    Weiler, M., Cesa, G.: General E(2)-equivariant steerable CNNs. In: NeurIPS (2019)

    2019

  23. [23]

    In: ACMMM

    Xie, J., Anwer, R.M., Cholakkal, H., Nie, J., Cao, J., Laaksonen, J., Khan, F.S.: Learning a dynamic cross-modal network for multispectral pedestrian detection. In: ACMMM. pp. 4043–4052 (2022)

    2022

  24. [24]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Xiong, Y., Li, Z., Chen, Y., Wang, W., Zhu, X., Zhang, L., Lu, T., Dai, J., Qiao, Y.: Efficient deformable convnets: Rethinking dynamic and sparse operator for vision applications. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 5765–5775 (2024).https://doi.org/10. 1109/CVPR52733.2024.00551

    arXiv 2024

  25. [25]

    IEEE Geoscience and Remote Sensing Letters (2024)

    Xu, F., Xu, T., Hong, L., Peng, P., Guo, J., Li, J.: Enhanced spectral-spatial fusion network for multispectral object detection in ground-aerial images. IEEE Geoscience and Remote Sensing Letters (2024)

    2024

  26. [26]

    In: CVPR

    Zhang, H., Wang, Y., Dayoub, F., Sunderhauf, N.: VarifocalNet: An IoU-aware dense object detector. In: CVPR. pp. 8514–8523 (2021)

    2021

  27. [27]

    Applied Soft Com- puting163, 111971 (2024)

    Zhang, J., Chang, T., Zhao, L., Ma, J., Han, B., Zhang, L.: Efficient cross-modality feature interaction for multispectral armored vehicle detection. Applied Soft Com- puting163, 111971 (2024)

    2024

  28. [28]

    In: Asian Conference on Computer Vision

    Zhang, Q., Qiu, L., Zhou, L., An, J.: ESM-YOLO: Enhanced small target detec- tion based on visible and infrared multi-modal fusion. In: Asian Conference on Computer Vision. pp. 1454–1469 (2024)

    2024

  29. [29]

    arXiv preprint arXiv:2401.10731 (2024)

    Zhao, T., Yuan, M., Jiang, F., Wang, N., Wei, X.: Removal and selection: Im- proving RGB-infrared object detection via coarse-to-fine fusion. arXiv preprint arXiv:2401.10731 (2024)

    arXiv 2024

  30. [30]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Zhu, H., Dong, W., Yang, L., Li, H., Yang, Y., Ren, Y., Zhu, Q., Feng, Z., Li, C., Lin, S., et al.: WaveMamba: Wavelet-driven mamba fusion for RGB-infrared object detection. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 11219–11229 (2025)

    2025

  31. [31]

    IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing18, 2401–2413 (2025)

    Zhu, J., Zhang, H., Li, S., Wang, S., Ma, H.: Cross teaching-enhanced multispectral remote sensing object detection with transformer. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing18, 2401–2413 (2025)

    2025

  32. [32]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019)

    Zhu, X., Hu, H., Lin, S., Dai, J.: Deformable convnets v2: More deformable, better results. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (June 2019)

    2019