FSDC-DETR: Frequency-Spatial Collaborative Detector

• The paper introduces FSDC-DETR, a framework that fuses spatial and frequency representations to boost small object detection accuracy.
• Its architecture incorporates a dual-branch backbone with DBFSAF, SFS-FF, and FSD-Down modules, ensuring high-frequency details are preserved.
• Experimental results on VisDrone-DET2019 and AITODv2 show gains of around 6.4–6.9 AP over DEIMv2-L, highlighting the method’s effectiveness.

Frequency-Spatial Domain Collaborative Detection Transformer (FSDC-DETR) is a small object detection framework that explicitly models complementary spatial and frequency representations within a DETR-style pipeline. It is presented as an enhanced variant of DEIMv2 for small object detection, with three principal modules—Dual-Branch Frequency-Spatial Adaptive Fusion (DBFSAF), Shunt Frequency-Spatial Feature Fusion (SFS-FF), and Frequency-Spatial Dynamic Downsampling (FSD-Down)—and is evaluated on VisDrone-DET2019 and AITODv2, where it reports improvements of 6.4 AP and 6.6 AP over DEIMv2-L, together with gains of 6.8 and 6.9 AP for small objects (Liu et al., 6 Jul 2026).

1. Definition and problem setting

FSDC-DETR addresses small object detection (SOD), particularly in aerial and UAV-style imagery, where objects occupy very few pixels and high-frequency cues are easily destroyed by ordinary feature processing. The motivating claim is that existing detectors remain limited by rigid processing that entangles spatial aggregation with implicit frequency aliasing and truncation, leading to inadequate preservation of high-frequency components for SOD. In this formulation, SOD is treated not only as a spatial-resolution problem but also as a frequency-preservation problem (Liu et al., 6 Jul 2026).

The two benchmarks emphasized in the method description illustrate this setting. VisDrone-DET2019 contains UAV images with crowded scenes and many small instances. AITODv2 is described as an especially hard tiny-object benchmark, with an average object size of 12.7 pixels and 86% of objects smaller than 16 pixels. The method therefore targets regimes in which local contrast, object boundaries, and fine structural patterns are critical to recall and localization quality (Liu et al., 6 Jul 2026).

A central premise of the architecture is that small objects depend heavily on high-frequency information, while conventional multi-scale fusion and downsampling attenuate that information. The paper characterizes CNNs as stronger in high-frequency and local-detail retention, and ViTs as stronger in low-frequency and global semantic modeling. FSDC-DETR is organized around exploiting that complementarity rather than collapsing both representations through simple concatenation and projection.

2. Architectural foundation

FSDC-DETR keeps the broad DEIMv2 structure but replaces its naive dual-branch fusion and standard hybrid encoder operations with explicitly frequency-aware mechanisms. The detector uses a dual-branch backbone combining HGNetv2 and a DINOv3-pretrained ViT. For stages $P \in \{2,3,4,5\}$, corresponding branch features are passed to DBFSAF, which produces refined stage features $F^P$. These are then processed by a Frequency-Spatial Collaborative Hybrid Encoder built from SFS-FF and FSD-Down, after which the encoded features are propagated to the decoder and standard DETR-style heads (Liu et al., 6 Jul 2026).

The pipeline can be summarized as follows in the paper’s own organization: input image, dual-branch backbone, per-stage DBFSAF, hybrid encoder with SFS-FF and FSD-Down, decoder, and final class and box prediction. The main value is not a new decoder, but a frequency-aware backbone and encoder redesign. This is an important architectural distinction: the novelty is concentrated in representation construction, cross-scale propagation, and scale-transition preservation rather than in Hungarian matching or decoder-layer reformulation.

Three modules structure that redesign. DBFSAF constructs frequency-spatial stage features from CNN and ViT outputs. SFS-FF performs cross-scale feature fusion through split-flow frequency, spatial, and residual paths. FSD-Down replaces conventional downsampling with a wavelet-based and grouped-convolution-based mechanism intended to preserve informative high-frequency responses during scale transitions. The coordination of these modules is what the paper identifies as “frequency-spatial collaborative.”

3. Dual-Branch Frequency-Spatial Adaptive Fusion

DBFSAF is the first explicit collaborative module. It is inserted after the dual-branch backbone outputs at stages $P \in \{2,3,4,5\}$, where $F_L^P$ and $F_H^P$ denote the two branch features. The module first applies Frequency Dynamic Convolution and then performs a partial channel split controlled by ratio $\gamma$:

$$F_C^P,\; F_{FS}^P = \operatorname{Partial}_{\gamma}\left(\operatorname{FDConv}\left(\operatorname{Concat}[F_L^P, F_H^P]\right)\right).$$

Here, $F_{FS}^P$ is the subset of channels subjected to full frequency-spatial refinement, while $F_C^P$ is preserved with lighter processing (Liu et al., 6 Jul 2026).

Two refinement branches are then constructed from $F_{FS}^P$:

$F^P$0

$F^P$1

The spatial branch uses the Multi-Kernel Spatial Refine Unit (MKSRU), while the frequency branch uses the Frequency Processing Unit (FPU). Their outputs are adaptively reweighted through global average pooling and softmax:

$F^P$2

$F^P$3

The final fused stage feature is

$F^P$4

The module’s empirical role is prominent. In the ablation study on VisDrone-DET2019, adding DBFSAF alone raises AP from 24.7 to 29.7 and $F^P$5 from 14.2 to 19.8, making it the strongest individual contributor among the three principal modules. The partial refinement ratio also matters: the best result is reported at $F^P$6, which suggests that refining only part of the channel space is preferable to either underusing or overextending the dual-domain interaction (Liu et al., 6 Jul 2026).

4. Collaborative hybrid encoder and scale-transition preservation

The hybrid encoder implements the paper’s second and third collaborative mechanisms. SFS-FF performs structure-aware cross-scale aggregation by receiving a current-scale feature $F^P$7 and a higher-resolution feature $F^P$8. After downsampling $F^P$9 with FSD-Down and concatenating it with $P \in \{2,3,4,5\}$0, the features are split into three paths:

$P \in \{2,3,4,5\}$1

The paper allocates $P \in \{2,3,4,5\}$2, $P \in \{2,3,4,5\}$3, and $P \in \{2,3,4,5\}$4 channels to these paths. Despite the notation, the next step routes $P \in \{2,3,4,5\}$5 through FFT/IFFT processing and $P \in \{2,3,4,5\}$6 through the spatial refine module:

$P \in \{2,3,4,5\}$7

This split-flow design is what the paper calls “shunt” frequency-spatial fusion (Liu et al., 6 Jul 2026).

FSD-Down is the dedicated scale-transition operator. Conventional strided convolution or pooling is treated as harmful because it causes implicit frequency aliasing and suppresses high-frequency responses. Instead, FSD-Down combines a DWT branch and a grouped-convolution branch. Channel attention is formed from both:

$P \in \{2,3,4,5\}$8

and the downsampled output is

$P \in \{2,3,4,5\}$9

The DWT produces the four sub-bands $F_L^P$0, preserving explicit low- and high-frequency components during downsampling (Liu et al., 6 Jul 2026).

The encoder therefore performs collaboration in two senses. First, it propagates frequency-enriched and structure-aware features across scales. Second, it avoids treating downsampling as a purely spatial resize operation. In ablations, SFS-FF alone improves AP from 24.7 to 25.9 and $F_L^P$1 from 14.2 to 15.5, while FSD-Down alone improves AP to 25.4 and $F_L^P$2 to 15.1. The combined model with DBFSAF, SFS-FF, and FSD-Down reaches 31.1 AP and 21.0 $F_L^P$3 on VisDrone-DET2019 (Liu et al., 6 Jul 2026).

5. Training protocol, benchmarks, and empirical performance

FSDC-DETR is trained for 100 epochs with batch size 64 and input size $F_L^P$4, using PyTorch 2.5.1 on 8 NVIDIA H200 GPUs. The optimizer is AdamW with learning rate $F_L^P$5, momentum 0.9, and weight decay $F_L^P$6. Following DEIMv2, the training recipe includes color jitter and zoom-out, and additionally MixUp, Mosaic, and CopyBlend. Dense O2O from DEIM is disabled after the first 50% of training, and all data augmentation is turned off in the final two epochs (Liu et al., 6 Jul 2026).

On VisDrone-DET2019, FSDC-DETR reports

$F_L^P$7

Compared with DEIMv2-L, this corresponds to $F_L^P$8 AP and $F_L^P$9 $F_H^P$0. On AITODv2, the reported results are

$F_H^P$1

which improve over DEIMv2-L by $F_H^P$2 AP and $F_H^P$3 $F_H^P$4. The model size reported in the main result table is 40.3M parameters (Liu et al., 6 Jul 2026).

The ablation study clarifies module contributions. The baseline without DBFSAF, SFS-FF, or FSD-Down gives AP $F_H^P$5 and $F_H^P$6. Adding DBFSAF only yields $F_H^P$7, adding SFS-FF only yields $F_H^P$8, adding FSD-Down only yields $F_H^P$9, and the full model reaches $\gamma$0. This pattern is consistent with the paper’s interpretation that the method succeeds by improving representation construction, cross-scale propagation, and scale-transition preservation rather than by modifying the decoder objective.

6. Relation to adjacent DETR research

FSDC-DETR belongs to a broader 2025–2026 line of DETR-style small-object detectors that increasingly treat frequency information as a first-class design variable. A closely related neighboring method is FSDETR, a frequency-spatial feature enhancement framework built on RT-DETR that uses SHAB, DA-AIFI, and FSFPN/CFSB, and reports 13.9% $\gamma$1 on VisDrone 2019 with 14.7M parameters (Huang et al., 16 Apr 2026). Another nearby design is $\gamma$2-DETR, which fuses spatial and frequency domain information to refine low-level features and improve density-guided tiny-object localization on AI-TOD-v2, raising AP from 28.7 to 31.3 relative to Dome-DETR with the same HGNetv2-B0 backbone (Wen et al., 6 Jan 2026).

In UAV detection, HEGS-DETR provides a complementary line of evidence by emphasizing high-frequency enhanced semantics, geometry-aware positional encoding, and efficient small-object pyramid design, reporting a 5.1% $\gamma$3 and 3.8% AP increase over its RT-DETR-based baseline while reducing parameter count by 4M (Peng et al., 1 Jul 2025). At a more foundational level, Deformable DETR established sparse reference-point-based multi-scale attention as a strong spatial alternative to dense quadratic attention, improving small-object detection and reducing training epochs by a factor of ten relative to DETR on COCO (Zhu et al., 2020).

Within that landscape, FSDC-DETR is distinguished by its explicit claim that standard detection operations entangle spatial aggregation with implicit frequency aliasing and truncation. Its contribution is therefore not merely to add a spectral branch, but to organize the detector around three coupled operations: construction of frequency-aware dual-branch features, progressive cross-scale propagation of spatial and frequency representations, and frequency-preserving scale transition. A common misconception is to interpret it as a decoder innovation; the paper instead presents it as a frequency-aware redesign of the backbone and hybrid encoder, with code released at github.com/nevereverinsomnia/FSDC-DETR (Liu et al., 6 Jul 2026).