FMC-DETR: Frequency-Decoupled Multi-Domain Coordination for Aerial-View Object Detection

Ben Liang; Hongguang Wei; Yuan Liu; Bingwen Qiu; Yihong Wang; Xiubao Sui; Qian Chen

cs.CVSep 27, 2025

FMC-DETR: Frequency-Decoupled Multi-Domain Coordination for Aerial-View Object Detection

Authors: Ben Liang, Hongguang Wei, Yuan Liu, Bingwen Qiu, Yihong Wang, Xiubao Sui, Qian Chen

Organizations: School of Electronic Engineering and Optoelectronic Technology, Nanjing University of Science and Technology, Nanjing 210094, China · State Key Laboratory of Extreme Environment Optoelectronic Dynamic Measurement Technology and Instrument, Nanjing 210094, China

Abstract

Remote sensing object detection is a critical technology for real-world applications such as natural resource monitoring, traffic management, and UAV-based rescue. Detecting tiny objects in high-resolution aerial imagery remains challenging due to weak visual cues and insufficient global context modeling in complex scenes. Existing methods often suffer from delayed contextual interaction and limited nonlinear reasoning, which restrict their ability to effectively refine shallow representations and ultimately lead to suboptimal performance. To address these challenges, we propose FMC-DETR, a frequency-decoupled fusion framework for aerial-view object detection. First, we propose the Wavelet Kolmogorov-Arnold Transformer (WeKat) backbone, which employs cascaded wavelet transforms to enhance global low-frequency structure perception in shallow features while preserving fine-grained details, and further leverages Kolmogorov-Arnold networks for adaptive nonlinear modeling of multi-scale dependencies. Second, we introduce the Multi-Domain Feature Coordination (MDFC) module, which refines cross-scale fused representations through partial-channel spatial, spectral, and structural coordination, thereby strengthening small-object-related feature responses in cluttered scenes. Finally, we design the Compact Partial Fusion (CPF) module, which performs compact multi-branch aggregation with progressive partial refinement to improve feature diversity and multi-scale interaction while preserving stable information flow and reducing redundant perturbation. Extensive experiments across multiple remote sensing benchmarks demonstrate that FMC-DETR achieves state-of-the-art performance and significantly outperforming the baseline detector. Code is available at https://github.com/bloomingvision/FMC-DETR.

Figures & tables

Fig. 1: Performance comparison between FMC-DETR-B and the Baseline [ 17 ] on the VisDrone, HazyDet, and SIMD datasets in terms of AP 50 , AP 75 , AP, AP S , and model parameters. The outer regions indicate better performance.

Fig. 2: Overview of the proposed FMC-DETR. The backbone consists of multiple stages to extract and model multi-scale features. Among them, HSG-WAVE and HSG-AKAT are applied to capture correlations in shallow and deep features, respectively. The multi-scale features are then fed into the CCFF structure for feature fusion. Finally, the Decoder and Detection Head query the optimized features to produce the final detection results.

Fig. 3: Overview of the proposed MDFC and CPF modules. The upper part shows the MDFC module, which consists of two successive phases: (1) amplitude-guided detail downsampling and (2) multi-domain refinement. The lower part illustrates the Compact Partial Fusion (CPF) module.

CNN-Based
Method	Publication	Input size	AP	AP 50	AP S	Param	FLOPs
YOLOv8-L [ 26 ]	Ultralytics	640 $\times$ 640	28.4	45.9	–	43.7M	165.2G
YOLOv8-X [ 26 ]	Ultralytics	640 $\times$ 640	28.9	46.8	–	68.2M	257.8G
YOLOv9-M [ 27 ]	ECCV2024	640 $\times$ 640	25.1	41.9	–	20.0M	76.3G
YOLOv10-L [ 28 ]	NeurIPS2024	640 $\times$ 640	27.6	44.6	–	24.4M	120.3G
YOLOv10-X [ 28 ]	NeurIPS2024	640 $\times$ 640	28.7	46.1	–	29.5M	160.4G

TABLE I: Quantitative comparison with state-of-the-art detectors on the VisDrone validation set. FMC-DETR-B (FMC-DETR-Base) denotes the default general configuration, and FMC-DETR-T (FMC-DETR-TinyAware) denotes the additional small-object-oriented variant. Results marked with † are reproduced by us under the same experimental settings. The best and second-best results are highlighted in red and blue , respectively.

Model	AP	AP car	AP truck	AP bus	Param
RT-DETR-R18	53.3	62.4	36.3	61.2	20.0M
IAYOLO	38.3	44.1	22.2	48.6	61.8M
MS-DAYOLO	48.3	59.4	28.5	57.0	40.0M
TOOD	51.4	58.4	33.6	62.2	32.0M
Cascade RCNN	51.6	59.0	34.2	61.7	69.1M
Deform DETR	51.5	58.4	33.9	62.3	40.0M

TABLE II: Comparison of the performance of different SOTA detectors on the HazyDet dataset.

Model	AP	AP 50	Param	FLOPs
RT-DETR-R18	63.7	78.6	20.0M	60.0G
YOLOv8-L	63.1	78.1	43.7M	165.2G
YOLOv9-M	62.2	76.6	20.0M	76.3G
DEIM-D-FINE-S	66.2	82.0	10.1M	24.9G
FBRT-YOLO-X	66.2	83.0	22.8M	185.8G
Deform DETR	59.7	75.6	40.0M	173.0G

TABLE III: Comparison of the performance of different SOTA detectors on the SIMD dataset.

WeKat	MDFC	CPF	AP	AP 50	AP S	Param	FLOPs
✗	✗	✗	26.7	44.6	18.5	20.0M	60.0G
✓	✗	✗	27.8	46.1	19.5	16.1M	53.3G
✗	✓	✗	28.2	46.7	19.8	20.2M	61.4G
✗	✗	✓	27.6	45.4	19.3	20.7M	61.2G
✓	✓	✗	28.9	47.6	20.3	16.6M	58.7G
✓	✓	✓	29.4	48.4	21.4	17.4M	62.9G

TABLE IV: Ablation studies on the VisDrone dataset.

WeKat	MDFC	CPF	HazyDet			SIMD
WeKat	MDFC	CPF	AP	AP 50	AP 75	AP	AP 50	AP 75
✗	✗	✗	53.3	73.6	62.2	63.7	78.6	72.4
✓	✗	✗	54.0	73.8	62.7	63.9	78.5	73.8
✗	✓	✗	54.7	73.7	63.1	64.6	79.2	74.3
✗	✗	✓	54.3	73.6	62.9	65.2	80.1	74.2
✓	✓	✗	54.8	74.2	63.5	64.4	78.7	75.3

TABLE V: Ablation studies on the HazyDet and SIMD datasets.

HSG-WAVE	HSG-AKAT	AP	AP 50	AP 75	Params	GFLOPs
✓	Asym.+GKAN	27.8	46.1	28.2	16.1	53.3
✓	Asym.+MLP	27.5	46.2	27.8	16.1	53.3
✓	Attn+MLP	27.1	46.0	27.1	16.3	53.6
✓	✗	27.4	45.9	27.4	16.0	53.0
✗	✗	27.3	46.0	27.3	16.3	59.0

TABLE VI: Progressive ablation of WeKat on the VisDrone validation set. Asym. and GKAN denote asymmetric attention and Group KAN, respectively. A ✗ indicates that the corresponding HSG module is replaced with a conventional C2f [ 26 ] feature-extraction block.

Spectral block	AP	AP 50	AP 75	Param	FLOPs
WAVE (Ours)	27.8	46.1	28.2	16.1M	53.3G
MDFC	27.7	45.9	27.9	16.3M	55.3G

TABLE VII: Ablation on the spectral modeling block in WeKat. We replace the wavelet-based branch in HSG-WAVE with the FFT-based spectral modeling used in MDFC.

$\alpha$	$e$	AP	AP 50	AP 75	AP S	FLOPs
Varying $\alpha$ with fixed $e=1.0$
0.25	1.00	26.7	44.6	26.9	19.0	42.0G
$0.50^{\dagger}$	$1.00^{\dagger}$	27.8	46.1	28.2	19.5	53.3G
0.75	1.00	28.2	46.8	28.5	20.1	69.9G
Varying $e$ with fixed $\alpha=0.50$
0.50	0.75	27.6	46.2	28.0	19.6	51.9G

TABLE VIII: Sensitivity analysis of the HSG design in WeKat.

\alpha

denotes the initial split ratio, and

e

denotes the Com-Stream allocation ratio.

\dagger

indicates the default setting.

Fig. 4: Comparison of validation loss curves between the baseline RT-DETR and the proposed WeKat.

Spa.	Edge.	Freq.	AP	AP 50	AP 75	AP S	AP M	AP L
✓	✗	✗	27.2	45.9	27.2	19.2	37.1	41.6
✓	✓	✗	28.0	46.3	28.5	19.7	38.4	42.6
✓	✗	✓	28.0	46.2	28.5	19.4	38.7	42.6
✗	✓	✓	27.7	46.2	28.1	19.5	38.1	43.3
✓	✓	✓	28.2	46.7	28.8	19.8	38.8	44.0

TABLE IX: Fine-grained ablation of the spatial, edge-aware, and frequency branches in the multi-domain refinement stage of MDFC on the VisDrone dataset.

Model	S5	Detect Layer	AP	AP 50	AP 75	AP S	Param	FLOPs
Baseline	✓	$[D3,D4,D5]$	26.7	44.6	26.9	18.5	20.0M	60.0G
WeKat	✓	$[D3,D4,D5]$	27.8	46.1	28.2	19.5	16.1M	53.3G
WeKat	✗	$[D2,D3,D4]$	30.9	50.2	31.8	23.5	12.9M	120.1G
WeKat	✗	$[D3,D4]$	30.8	50.2	31.5	23.3	12.8M	103.0G
WeKat	✗	$[D2,D4]$	31.6	50.9	32.9	23.9	12.8M	115.8G
FMC-DETR-B	✓	$[D3,D4,D5]$	29.4	48.4	30.2	21.4	17.4M	62.9G

TABLE X: Ablation Study on the Effect of

S5

Backbone Layer and Detection Head Design. “✓” and “✗” indicate the presence or absence of the

S5

stage, respectively.

Fig. 5: Visualization of feature activations across detection layers for small objects. Here, $[D2-D5]$ denote detection layers at different feature scales, where the spatial resolution decreases progressively by a factor of 2 at each stage.

Fig. 6: Precision–Recall curves of the baseline RT-DETR on the VisDrone dataset at IoU=0.50.

Fig. 7: Precision–Recall curves of the proposed FMC-DETR-B on the VisDrone dataset at IoU=0.50.

Activation Function	AP	AP 75	AP 50	AP S
ReLU	27.3	27.6	45.2	19.4
SiLU	27.8	28.2	46.1	19.5

TABLE XI: Comparison of activation functions on the VisDrone dataset.

Method	Model size	Latency (ms/img)	FPS (img/s)
RT-DETR-R18	77.0 MB	15.4	64.8
FMC-DETR-B	67.1 MB	22.7	44.0
FMC-DETR-T	53.4 MB	20.7	48.4

TABLE XII: Inference efficiency under batch size 1. Latency and FPS are averaged over five runs on an NVIDIA RTX 4090D GPU with an input size of

640\times 640

Fig. 8: Visualizations of the detection results of baseline and our proposed FMC-DETR-B on HazyDet.

Fig. 9: Representative failure cases on the VisDrone validation set.

Fig. 10: Qualitative comparison of detection results produced by different methods across three aerial-view datasets. Each column corresponds to a different dataset, and the last row shows the results of the proposed framework, which exhibits more accurate localization and fewer missed or false detections.

Aerial Object Detection

Aerial object detection locates and classifies objects in drone and satellite imagery, often under scale, orientation, clutter, and compute constraints.

See all 121 papers

Apr 20, 2026

HMR-Net: Hierarchical Modular Routing for Cross-Domain Object Detection in Aerial Images

Despite advances in object detection, aerial imagery remains a challenging domain, as models often fail to generalize across variations in spatial resolution, scene composition, and semantic label coverage. Differences in geographic context, sensor characteristics, and object distributions across datasets limit the capacity of conventional models to learn consistent and transferable representations. Shared methods trained on such data tend to impose a unified representation across fundamentally different domains, resulting in poor performance on region-specific content and less flexibility when dealing with novel object categories. To address this, we propose a novel modular learning framework that enables structured specialization in aerial detection. Our method introduces a hierarchical routing mechanism with two levels of modularity: a domain routing layer that uses latent geographic embeddings to assign inputs to domain-specialized expert modules, and a scene routing mechanism that allocates image subregions to scene-specific expert modules. This allows our method to specialize across datasets and within complex scenes. Additionally, the framework contains a conditional expert module that uses external semantic information (e.g., category names or textual descriptions) to enable detection of novel object categories during inference, without the need for retraining or fine-tuning. By moving beyond monolithic representations, our method provides an adaptive framework for remote sensing object detection. Comprehensive evaluations on four datasets highlight improvements in multi-dataset generalization, region-level specialization, and open-category detection.

Pourya Shamsolmoali, Masoumeh Zareapoor, Michael Felsberg +3

Department of Computer Science, University of York, UK · SEIEE, Shanghai Jiao Tong University, China · Computer Vision Laboratory, Linkoping University, Sweden +2

Open-Vocabulary Object Detection Adaptive Model Routing Mixture of Experts Cross-Domain Transfer Learning Object Detection Aerial Object Detection

Ask PDF

Oct 7, 2026

LiG-DETR: Local-in-Global Reassembly in Latent Space for Aerial Object Detection

Aerial object detection faces substantial scale and density variations. Small objects are easily degraded by downsampling and feature compression, while medium and large objects require sufficient global context. Existing methods mainly follow two paradigms: image slicing provides clearer local evidence but relies on independent crop-level prediction and post-processing, whereas feature- and query-level optimization preserves unified inference but operates on already compressed full-image representations, limiting recovery of fine-grained information. This raises a key question: can aerial detection directly acquire high-fidelity local evidence before feature degradation and integrate it into a unified end-to-end framework? To this end, we propose LiG-DETR, an Efficient Global-Local Reassembly framework that reformulates image slicing as high-fidelity local feature acquisition. A shared encoder extracts global and locally magnified features, which are projected into the detector feature space. The projected local features are reassembled according to their original spatial locations to form a globally aligned local feature level, and a single DETR decoder jointly decodes global and local features. To reduce redundant computation, Context-Preserved Selective Reassembly focuses high-resolution encoding on informative regions while preserving a dense feature layout, and Density-Aware Adaptive Query Allocation adapts the decoder query budget using encoder proposal scores. Experiments show substantial gains on small and medium objects while retaining strong large-object performance, with favorable accuracy--efficiency trade-offs and improved cross-domain generalization. The code will be released.

Yupeng Zhang, Fangzhuo Gao, Juntao Cheng +3

Tianjin University. · Shanghai Jiaotong University. · Shanghai AI Laboratory. +1

Aerial Object Detection Remote Sensing Small Object Detection Efficient Object Detection Detection Transformer

Ask PDF

Apr 3, 2026

SFFNet: Synergistic Feature Fusion Network With Dual-Domain Edge Enhancement for UAV Image Object Detection

Object detection in unmanned aerial vehicle (UAV) images remains a highly challenging task, primarily caused by the complexity of background noise and the imbalance of target scales. Traditional methods easily struggle to effectively separate objects from intricate backgrounds and fail to fully leverage the rich multi-scale information contained within images. To address these issues, we have developed a synergistic feature fusion network (SFFNet) with dual-domain edge enhancement specifically tailored for object detection in UAV images. Firstly, the multi-scale dynamic dual-domain coupling (MDDC) module is designed. This component introduces a dual-driven edge extraction architecture that operates in both the frequency and spatial domains, enabling effective decoupling of multi-scale object edges from background noise. Secondly, to further enhance the representation capability of the model's neck in terms of both geometric and semantic information, a synergistic feature pyramid network (SFPN) is proposed. SFPN leverages linear deformable convolutions to adaptively capture irregular object shapes and establishes long-range contextual associations around targets through the designed wide-area perception module (WPM). Moreover, to adapt to the various applications or resource-constrained scenarios, six detectors of different scales (N/S/M/B/L/X) are designed. Experiments on two challenging aerial datasets (VisDrone and UAVDT) demonstrate the outstanding performance of SFFNet-X, achieving 36.8 AP and 20.6 AP, respectively. The lightweight models (N/S) also maintain a balance between detection accuracy and parameter efficiency. The code will be available at https://github.com/CQNU-ZhangLab/SFFNet.

Wenfeng Zhang, Jun Ni, Yue Meng +4

College of Computer and Information Science, Chongqing Normal University, Chongqing 401333, China · CETC Yizhihang (Chongqing) Technology Co., Ltd, Chongqing 400031, China · School of Computer Engineering, Weifang University, Weifang 261061, China +1

Object Detection Aerial Object Detection Efficient Object Detection Multi-Scale Feature Fusion

Ask PDF

Abstract

Aerial Object Detection

Explore similar work

HMR-Net: Hierarchical Modular Routing for Cross-Domain Object Detection in Aerial Images

LiG-DETR: Local-in-Global Reassembly in Latent Space for Aerial Object Detection

SFFNet: Synergistic Feature Fusion Network With Dual-Domain Edge Enhancement for UAV Image Object Detection