UAV-DETR: Efficient Drone Detection Framework

• The paper's main contribution is the integration of MSFF-FE, FD, and SAC modules into the DETR pipeline to preserve high-frequency details in UAV imagery.
• UAV-DETR builds on RT-DETR with modifications focused on multi-scale feature fusion and semantic alignment to address challenges like scale variation and occlusion.
• Empirical results show significant improvements in AP and AP50 on VisDrone and UAVVaste benchmarks while maintaining similar parameter counts to RT-DETR.

UAV-DETR denotes, in its strictest sense, the RT-DETR-derived framework introduced for unmanned aerial vehicle imagery in early 2025, and, in a broader later usage, a wider class of DETR-style detectors specialized for UAV, drone, and anti-drone detection. The specific 2025 formulation keeps RT-DETR’s end-to-end set-prediction pipeline while inserting UAV-oriented feature modules for frequency-aware multi-scale fusion, down-sampling, and semantic alignment, and subsequent papers reuse the label or treat it as a shorthand for UAV-specialized DETR research more generally (Zhang et al., 3 Jan 2025, Stefanescu et al., 11 Sep 2025, Yang et al., 24 Mar 2026).

1. Terminology and scope

A recurrent source of ambiguity is that “UAV-DETR” is both a paper title and a broader naming convention. In the specific paper "UAV-DETR: Efficient End-to-End Object Detection for Unmanned Aerial Vehicle Imagery" (Zhang et al., 3 Jan 2025), the term refers to a real-time, end-to-end detector tailored to UAV imagery through three added modules: Multi-Scale Feature Fusion with Frequency Enhancement (MSFF-FE), Frequency-focused Down-sampling (FD), and Semantic Alignment and Calibration (SAC), together with Inner-SIoU for regression. In later literature, the expression is also used more generically: WAVE-DETR explicitly states that many practitioners use “UAV-DETR” as shorthand for DETR-style detectors applied to UAV or drone detection, and a distinct 2026 anti-drone detector is itself titled "UAV-DETR: DETR for Anti-Drone Target Detection" (Stefanescu et al., 11 Sep 2025, Yang et al., 24 Mar 2026).

Across these usages, the common conceptual core is DETR-style end-to-end detection: anchor-free prediction, one-to-one Hungarian matching, and elimination of NMS. The UAV-specific motivation is also stable across the literature: UAV imagery is characterized by small targets, dense distributions, scale variation, occlusion, cluttered backgrounds, and real-time or edge-compute constraints (Zhang et al., 3 Jan 2025, Chen, 24 Mar 2025, Wenbin, 21 Apr 2025). This suggests that UAV-DETR is best understood not as a single immutable architecture, but as a research lineage centered on adapting DETR-family detectors to aerial imaging regimes where small-object fidelity and cross-scale consistency dominate design choices.

2. Canonical 2025 architecture

The canonical 2025 UAV-DETR is built on RT-DETR and keeps its end-to-end set prediction pipeline and transformer-based detection head unchanged. The reported variants use ResNet-18 and ResNet-50 backbones, remain anchor-free and NMS-free, and inherit the transformer encoder/decoder design, object queries, and positional encoding from RT-DETR without structural changes to the transformer stack (Zhang et al., 3 Jan 2025). The architectural specialization is concentrated in feature extraction and fusion rather than in the DETR core.

Its three defining modules operate before decoding. MSFF-FE is a multi-scale feature fusion with frequency enhancement module intended to preserve and amplify high-frequency detail relevant to small objects. FD is a frequency-focused down-sampling operator that uses parallel dual-domain paths so that resolution reduction does not discard critical spatial detail. SAC is a semantic alignment and calibration module that aligns and fuses multi-path features using learned offsets and gating in order to reduce spatial and semantic misalignment introduced by dual-domain fusion (Zhang et al., 3 Jan 2025).

The result is a UAV-specific reformulation of RT-DETR in which the central architectural claim is not a new query mechanism or a new decoder, but a different treatment of feature transport across scales. The original paper states that classic spatial-only fusion and down-sampling progressively discard edges and textures that define small objects in UAV scenes, and the three added modules are explicitly designed to counter that failure mode (Zhang et al., 3 Jan 2025).

3. Frequency-aware feature processing and objective design

MSFF-FE uses a CSP-style split in which the input feature $x \in \mathbb{R}^{C \times H \times W}$ is divided into a branch with $C_1 = C/4$ channels and a bypass branch with $C_2 = 3C/4$ channels. The enhanced branch first applies a $1 \times 1$ convolution and GELU, then performs learned spectral gating in the Fourier domain:

$$x_{sp} = \left| \mathcal{F}^{-1}\!\left( H \odot \mathcal{F}(x_{conv}) \right) \right|$$

It then applies mixed receptive fields,

$$x_{sc} = Conv_{1\times1}(x_{sp}) + Conv_{3\times3}(x_{sp}) + Conv_{5\times5}(x_{sp}),$$

followed by a second dual-domain refinement,

$$x_F = \alpha \cdot \mathcal{F}^{-1}\!\left( \mathcal{F}(Conv_{1\times1}(x_{sc})) \odot Conv_{1\times1}(x_{sc}) \right) + \beta \cdot x_{sc},$$

and final branch fusion,

$$x_{final} = x_1 + Conv_{31\times31}(x_{conv}) + Conv_{1\times1}(x_{conv}) + x_F.$$

The module output is obtained by concatenating $x_{final}$ with the bypass branch and applying a final $1 \times 1$ convolution (Zhang et al., 3 Jan 2025).

FD performs pre-pooling by $C_1 = C/4$0, then splits channels into two paths. One path uses a stride-2 $C_1 = C/4$1 convolution, while the other combines a frequency branch and a pooled branch before a $C_1 = C/4$2 fusion. In compact form, the down-sampling operator is reported as

$C_1 = C/4$3

SAC then attempts to correct alignment errors created by dual-path fusion. After channel unification and resizing, it performs a gated spatial–spectral merge,

$C_1 = C/4$4

followed by learned geometric alignment,

$C_1 = C/4$5

and calibrated fusion,

$C_1 = C/4$6

This design explicitly treats frequency enhancement and spatial alignment as coupled operations rather than independent neck components (Zhang et al., 3 Jan 2025).

Set-based matching itself follows the RT-DETR recipe. The matching permutation is

$C_1 = C/4$7

with cost

$C_1 = C/4$8

The training loss modifies the regression term by replacing GIoU with Inner-SIoU:

$C_1 = C/4$9

The paper reports that Inner-IoU is computed on scaled inner boxes and that the best scaling ratio on VisDrone is $C_2 = 3C/4$0 (Zhang et al., 3 Jan 2025).

4. Training protocol, empirical results, and efficiency of the original model

The original UAV-DETR is evaluated on VisDrone-2019-DET and UAVVaste. The reported training setup uses image size $C_2 = 3C/4$1, RT-DETR data augmentation plus Mosaic with $C_2 = 3C/4$2 and MixUp with $C_2 = 3C/4$3, AdamW with base learning rate $C_2 = 3C/4$4, batch size $C_2 = 3C/4$5, $C_2 = 3C/4$6 epochs, early stopping with patience $C_2 = 3C/4$7, and an NVIDIA GeForce RTX 3090 (Zhang et al., 3 Jan 2025).

On VisDrone validation, the reported gains over RT-DETR are exact and substantial for both backbone scales. The model preserves the same parameter count as the corresponding baseline but increases GFLOPs because of the spectral modules.

Model	Params / GFLOPs / FPS	VisDrone AP / AP50
RT-DETR-R18	20 M / 60 / 85	26.7 / 44.6
UAV-DETR-R18	20 M / 77 / 51	29.8 / 48.8
RT-DETR-R50	42 M / 136 / 40	28.4 / 47.0
UAV-DETR-R50	42 M / 170 / 30	31.5 / 51.1

The reported improvements are +3.1 AP and +4.2 AP50 for the R18 configuration, and +3.1 AP and +4.1 AP50 for the R50 configuration (Zhang et al., 3 Jan 2025). On UAVVaste, RT-DETR reports AP 42.4 and AP50 71.1, while UAV-DETR reports AP 45.7 and AP50 74.7, corresponding to gains of +3.3 AP and +3.6 AP50 (Zhang et al., 3 Jan 2025).

The VisDrone ablation sequence localizes the contribution of each added component. Starting from RT-DETR-R18 at 26.7 AP and 44.6 AP50, adding Inner-SIoU yields 27.1 and 45.3; adding MSFF-FE yields 28.4 and 46.9; adding FD yields 28.4 and 47.1; adding SAC without FD yields 28.9 and 47.7; and the full model reaches 29.8 and 48.8 (Zhang et al., 3 Jan 2025). The paper also notes a qualitative limitation: attention occasionally focuses on irrelevant or noisy regions.

5. Subsequent UAV-specialized DETR variants

The literature that follows broadens the design space while preserving DETR-style end-to-end detection. Some variants remain closest to the original UAV-DETR’s emphasis on frequency-aware fusion, whereas others shift toward local–global interaction, variable receptive fields, sparse attention, or decoder stabilization. The papers summarized below are representative rather than exhaustive (Chen, 24 Mar 2025, Wenbin, 21 Apr 2025, Peng et al., 1 Jul 2025, Shufang, 11 Sep 2025, Gao et al., 8 Dec 2025, Chen et al., 26 Feb 2026).

Model	Main architectural emphasis	Reported result
LGI-DETR (Chen, 24 Mar 2025)	LSE and GII for bidirectional local–global interaction in the encoder	VisDrone: 28.1 AP, 46.0 AP50
VRF-DETR (Wenbin, 21 Apr 2025)	MSCF, GConv, and GMCF with variable receptive fields	VisDrone: 51.4% mAP50, 31.8% mAP50:95, 13.5M params
HEGS-DETR (Peng et al., 1 Jul 2025)	HFESNet, ESOP, GAPE, and SQR	VisDrone val: 48.0 AP50, 29.4 AP, 16.53M params
RT-DETR++ (Shufang, 11 Sep 2025)	AU, AD, and CSP-PAC in the encoder neck	VisDrone val: AP 24.1, AP_s 14.2, latency 5.5 ms
DFIR-DETR (Gao et al., 8 Dec 2025)	DCFA, DFPN, and FIRC3 with dynamic K-sparse attention and frequency refinement	VisDrone: 51.6% mAP50, 31.6% mAP50:95, 11.7M params
UFO-DETR (Chen et al., 26 Feb 2026)	LSKNet, DAttention-AIFI, and DynFreq-C3	VisDrone: mAP50 46.1, 41.8 GFLOPs, 28.3 MB

LGI-DETR keeps RT-DETR’s decoder and adds Local Spatial Enhancement at the encoder’s initial stage and Global Information Injection at its final stage, reporting +1.9 AP and +2.4 AP50 over RT-DETR-r18 on VisDrone and gains on UAVDT as well (Chen, 24 Mar 2025). VRF-DETR replaces RT-DETR’s AIFI with MSCF, upgrades backbone blocks to GMCF, and distributes GConv operators through the network, reporting 51.4% mAP50 and 31.8% mAP50:95 with only 13.5M parameters (Wenbin, 21 Apr 2025). HEGS-DETR augments RT-DETR-R18 with HFESNet, ESOP, Geometry-Aware Positional Encoding, and Selective Query Recollection, achieving +5.1 AP50 and +3.8 AP on VisDrone validation while reducing parameter count by 3.35–4M relative to the baseline (Peng et al., 1 Jul 2025).

RT-DETR++ is explicitly encoder-centric: it retains RT-DETR’s backbone and Transformer decoder and redesigns feature propagation through Attention Upsampling, Attention Downsampling, and CSP-PAC. On VisDrone validation, the full configuration reports AP 24.1, AP50 40.1, AP75 23.6, AP_s 14.2, and 5.5 ms latency on RTX 4090D, compared with 21.9 AP and 12.5 AP_s for RT-DETR-R18 (Shufang, 11 Sep 2025). DFIR-DETR combines dynamic K-sparse attention, amplitude-normalized upsampling, and frequency-domain refinement, reporting 51.6% mAP50, 31.6% mAP50:95, 11.7M parameters, 41.2 GFLOPs, and 40.1 FPS on VisDrone (Gao et al., 8 Dec 2025). UFO-DETR uses an LSKNet backbone, deformable attention inside AIFI, and DynFreq-C3 for frequency-enhanced decoding, reporting Precision 59.2, Recall 44.5, mAP50 46.1, 41.8 GFLOPs, and 28.3 MB on VisDrone2019 (Chen et al., 26 Feb 2026).

The literature surveyed here suggests three especially persistent research directions: preservation of high-frequency or frequency-domain detail, explicit multi-scale fusion for small targets, and computationally efficient attention or query refinement. Those directions are already latent in the original UAV-DETR and become progressively more explicit in later variants (Zhang et al., 3 Jan 2025, Peng et al., 1 Jul 2025, Gao et al., 8 Dec 2025).

6. Anti-drone, multimodal, and inspection-oriented reinterpretations

A second major development is expansion beyond generic aerial scene detection into specialized domains. The 2026 anti-drone model titled "UAV-DETR: DETR for Anti-Drone Target Detection" (Yang et al., 24 Mar 2026) is distinct from the 2025 UAV imagery detector. It introduces a WTConv-enhanced backbone, a Sliding Window Self-Attention encoder termed SWSA-IFI, an Efficient Cross-Scale Feature Recalibration and Fusion Network, and a hybrid Inner-CIoU plus NWD loss. On its custom UAV dataset, it reports a +6.61% improvement in mAP50:95 over RT-DETR with a 39.8% reduction in parameters, and on DUT-ANTI-UAV it reports +1.4% Precision and +1.0% F1-Score over RT-DETR (Yang et al., 24 Mar 2026). Here, the “UAV” in UAV-DETR refers not to imagery acquired from a drone platform, but to the drone as a target in counter-UAV detection.

WAVE-DETR moves in a different direction by making UAV detection multi-modal rather than purely visual. It fuses Wav2Vec2 acoustic embeddings with Deformable DETR multi-resolution visual features through gated, linear, MLP, and cross-attention fusion strategies. The best performer is the gated fusion approach, which improves small-drone mAP by 11.1% to 15.3% across IoU thresholds from 0.5 to 0.9 on ARDrone in-distribution and out-of-distribution settings, and reports overall gains across all drone sizes ranging from 3.27% to 5.84% (Stefanescu et al., 11 Sep 2025). This broadens the conceptual space of UAV-DETR from aerial-image object detection to synchronized audio–visual drone sensing.

TinyDef-DETR exemplifies yet another reinterpretation: UAV-based power-line inspection. Built on RE-DETR rather than RT-DETR, it replaces strided down-sampling with stride-free space-to-depth, inserts edge-enhanced convolution, adds a cross-stage dual-domain multi-scale attention module, and uses Focaler-Wise-SIoU for regression. On CSG-ADCD it reports Precision 0.534, Recall 0.263, mAP50 0.275, and mAP50:95 0.1187, versus 0.163 and 0.0738 mAP50 and mAP50:95 for RE-DETR-R18; on VisDrone it reports 0.372 mAP50 and 0.217 mAP50:95 (Cui, 7 Sep 2025). A plausible implication is that, once DETR is adapted to preserve tiny structures and boundary detail, the UAV-DETR paradigm transfers to narrow industrial inspection tasks as well as to generic aerial-object benchmarks.

7. Limitations and open research directions

Despite the diversity of proposed modules, the surveyed papers report a stable set of unresolved failure modes. The original UAV-DETR notes occasional attention on irrelevant or noisy regions (Zhang et al., 3 Jan 2025). RT-DETR++ explicitly reports difficulty with objects smaller than 4 pixels, heavily occluded scenes, and a modest increase in inference latency (Shufang, 11 Sep 2025). VRF-DETR identifies extreme density, motion blur, low light, and altitude or sensor shifts as persistent weaknesses (Wenbin, 21 Apr 2025). The anti-drone UAV-DETR reports biological distractors such as birds and severe camouflage against complex architectural textures as remaining failure cases, while also acknowledging increased FLOPs relative to RT-DETR (Yang et al., 24 Mar 2026).

Future work proposed across the literature is correspondingly consistent. RT-DETR++ proposes more lightweight contextual modeling and validation on more diverse UAV datasets and adverse conditions such as smog and atmospheric turbulence (Shufang, 11 Sep 2025). VRF-DETR proposes temporal modeling, domain adaptation, and sensor fusion (Wenbin, 21 Apr 2025). UFO-DETR points to redundant computational overhead in the positional relationship decoder within RT-DETR and emphasizes the need for clearer formalization of losses, queries, and positional encodings (Chen et al., 26 Feb 2026). WAVE-DETR motivates larger synchronized audio–visual datasets, better cross-modal alignment, and robustness to adversarial and camouflage conditions (Stefanescu et al., 11 Sep 2025). Taken together, this suggests that the long-term trajectory of UAV-DETR research is toward stronger tiny-object sensitivity, more reliable deployment under domain shift, and tighter integration of temporal, multimodal, and hardware-aware design constraints.