Synergistic Feature Pyramid Network (SFPN)

• SFPN is a specialized feature fusion module within SFFNet that preserves low-level spatial details and enhances small-object responses in UAV imagery.
• It employs Linear Deformable Convolution (LDConv) and a Wide-Area Perception Module (WPM) to adaptively capture geometric variability and long-range contextual cues.
• Empirical evaluations show that SFPN improves AP metrics significantly over traditional FPNs, demonstrating efficient multi-scale fusion without extra detection heads.

Searching arXiv for the relevant SFPN papers and acronym usage. The term Synergistic Feature Pyramid Network (SFPN) denotes the neck module introduced in SFFNet for UAV image object detection, where it performs multi-scale feature fusion and enhancement after backbone extraction and before parallel detection heads (Zhang et al., 3 Apr 2026). In that formulation, SFPN is designed to address two stated weaknesses of conventional pyramid-based fusion in aerial detection: the attenuation of small-object responses in deep feature maps and the underuse of low-level spatial detail and long-range context (Zhang et al., 3 Apr 2026). The acronym SFPN is also used in an earlier and distinct sense for Synthetic Fusion Pyramid Network, a modification of standard FPN for one-stage object detection that inserts synthetic intermediate pyramid levels to smooth scale transitions (Zhang et al., 2022). This acronym overlap is consequential for the literature: the 2026 usage emphasizes synergistic fusion with linear deformable convolution and a wide-area perception module, whereas the 2022 usage emphasizes denser scale sampling via synthetic layers (Zhang et al., 3 Apr 2026, Zhang et al., 2022).

1. Terminological scope and acronym disambiguation

In the SFFNet framework, SFPN is explicitly the neck situated between the backbone and the detection heads, and its purpose is to make the feature hierarchy more suitable for UAV object detection by preserving fine-grained low-level information, strengthening small-object representation, and injecting contextual awareness into high-resolution features (Zhang et al., 3 Apr 2026). The framework overview specifies a key pathway: the C2 layer features are first processed by linear deformable convolution and then forwarded to the P3 layer, where they are further enhanced by WPM (Zhang et al., 3 Apr 2026).

A separate paper, "SFPN: Synthetic FPN for Object Detection" (Zhang et al., 2022), uses the same acronym for Synthetic Fusion Pyramid Network. That method is presented as a modification of the standard Feature Pyramid Network (FPN) for one-stage object detection, motivated by the claim that adjacent FPN levels separated by a factor of 2 create a scale truncation problem and do not fuse features smoothly (Zhang et al., 2022). It introduces synthetic intermediate pyramid levels so that objects of nearby sizes are less likely to be assigned to mismatched prediction maps (Zhang et al., 2022).

This dual usage means that SFPN is not a uniquely identifying acronym in the arXiv literature. A plausible implication is that any technical discussion of “SFPN” requires explicit expansion of the term and citation context, especially because the two methods target related but non-identical deficiencies of conventional feature pyramids.

2. Position of the synergistic SFPN within SFFNet

SFFNet is organized into three parts: a backbone that extracts multi-scale features and is strengthened by the paper’s MDDC module, a neck in which SFPN performs multi-scale feature fusion and enhancement, and two parallel heads that decode class and box predictions, with the one-to-many head being training-only (Zhang et al., 3 Apr 2026). Within this pipeline, SFPN operates after backbone feature extraction and before prediction decoding.

The paper frames SFPN as a response to two issues in aerial detection. First, small objects lose detail in deep feature maps because feature responses are attenuated by downsampling. Second, traditional pyramid fusion underuses low-level details and long-range context, including spatial details from shallow layers, irregular object geometry, and contextual dependencies around the target (Zhang et al., 3 Apr 2026). The authors additionally note that directly adding extra low-level detection heads can improve detail usage but increases computational cost, and they present SFPN as a more efficient alternative based on feature fusion plus targeted enhancement modules rather than extra heads (Zhang et al., 3 Apr 2026).

The core design philosophy is described as a progressive, layer-by-layer synergistic fusion mechanism. It begins with high-level features, then integrates lower-level ones, uses a bottom-up fusion strategy, and employs concatenation so as to retain complete information without altering feature ratios (Zhang et al., 3 Apr 2026). This places the method conceptually within the broader family of pyramid necks, but with an explicit emphasis on information preservation during fusion rather than aggressive blending.

3. Internal structure: LDConv and WPM

The synergistic SFPN has two emphasized submodules: Linear Deformable Convolution (LDConv) on the P2/C2 low-level pathway and the Wide-Area Perception Module (WPM) at the P3 level (Zhang et al., 3 Apr 2026). The paper’s own summary assigns distinct roles to them: LDConv enhances fine detail extraction at the lowest pyramid level, while WPM captures both target detail and contextual information at a higher-resolution stage (Zhang et al., 3 Apr 2026).

LDConv

LDConv is motivated by the observation that UAV objects often undergo large appearance changes caused by viewpoint variation, distance changes, rotation, and scale variation (Zhang et al., 3 Apr 2026). The paper contrasts this with standard convolution, which uses a fixed kernel grid that is poorly matched to irregular geometric changes (Zhang et al., 3 Apr 2026). LDConv instead dynamically adjusts the sampling locations of the kernel, and the paper explicitly states that it can capture irregular object shapes, adaptively localize geometric structure, and do so with only a linear increase in parameter count (Zhang et al., 3 Apr 2026).

Its placement is also specific: LDConv is applied to the C2 feature map from the backbone, identified as the lowest-level input in the top-down path aggregation of SFPN (Zhang et al., 3 Apr 2026). Because C2 has the richest spatial detail and is especially useful for detecting small objects, applying LDConv there is intended to extract more useful information from shallow features before upward fusion (Zhang et al., 3 Apr 2026). The paper further characterizes this pathway as a feature-correlation pathway from low-level backbone features into the pyramid and states that it replaces the need for extra low-level detection heads while avoiding their computational burden (Zhang et al., 3 Apr 2026).

WPM

The Wide-Area Perception Module (WPM) is introduced because detection in UAV imagery depends not only on object appearance but also on surrounding context, particularly when objects are tiny, densely packed, elongated, or direction-sensitive (Zhang et al., 3 Apr 2026). The authors argue that small kernels have limited receptive fields and cannot model long-range dependencies well, whereas aerial scenes require both global context and anisotropic structure (Zhang et al., 3 Apr 2026).

WPM is placed at the P3 level because that level has relatively high resolution and still preserves enough spatial detail while being semantically stronger than lower layers (Zhang et al., 3 Apr 2026). Its architecture is described as a parallel composition of a large-kernel convolution $31\times31$, strip convolutions $1\times31$ and $31\times1$, a $1\times1$ convolution, and an identity path (Zhang et al., 3 Apr 2026). Only one quarter of the channels are sent into the parallel convolution branch to reduce computation (Zhang et al., 3 Apr 2026).

The stepwise flow is explicitly given. For input tensor

$X \in \mathbb{R}^{C \times H \times W},$

the module first processes $X$ through a CBS component, then splits channels into

$$X_1 \in \mathbb{R}^{\frac{C}{4}\times H\times W}, \qquad X_2 \in \mathbb{R}^{\frac{3C}{4}\times H\times W}.$$

Next, $X_1$ is passed through a $1\times1$ convolution and then through parallel depthwise convolutions $1\times1$, $1\times31$0, $1\times31$1, and $1\times31$2; an identity branch preserves original features; the branch outputs are concatenated along the channel dimension, fused with a final $1\times31$3 convolution, and then merged back with $1\times31$4 to form the output (Zhang et al., 3 Apr 2026).

The functional interpretation given in the paper is that the $1\times31$5 kernel captures isotropic global context, whereas the strip convolutions capture anisotropic / direction-sensitive context (Zhang et al., 3 Apr 2026). The strip convolutions are said to retain long-range dependence along a single axis while reducing irrelevant background interference from the orthogonal direction, which is particularly relevant for aerial targets with elongated or orientation-sensitive structure (Zhang et al., 3 Apr 2026).

4. Fusion mechanism and representational intent

The paper presents SFPN as improving feature fusion in three principal ways. First, by applying LDConv to low-level C2 features, it preserves fine-grained information and geometric irregularity that would otherwise be lost or suppressed in conventional pyramid fusion (Zhang et al., 3 Apr 2026). Second, by applying WPM at P3, it enhances contextual reasoning by expanding the receptive field and modeling wide-area dependencies (Zhang et al., 3 Apr 2026). Third, by using concatenation rather than more aggressive blending, it preserves the full information content of each feature level (Zhang et al., 3 Apr 2026).

The paper explicitly states that SFPN “preserves feature information from all layers,” “avoids the suppression of low-level features,” and “ensures that spatial information and details are fully transmitted” (Zhang et al., 3 Apr 2026). In this sense, the “synergistic” attribute refers not merely to cross-scale aggregation but to the joint preservation of shallow spatial detail, object-aligned geometric sampling, and wider contextual dependencies within a single neck design.

The target application is explicitly small-object detection in UAV imagery. The method is said to improve small-target detection by combining high-resolution detail preservation, shape-adaptive sampling, and contextual enhancement (Zhang et al., 3 Apr 2026). A plausible implication is that the neck is designed not only for scale fusion in the usual FPN sense, but for compensating for the particular pathologies of aerial scenes, including density, blur, and directional structure.

5. Empirical characterization in UAV object detection

The paper reports ablations that isolate the contribution of SFPN within SFFNet. In Table 4, adding SFPN to the baseline increases AP from 27.2% to 31.3%, while AP$1\times31$6 increases by 3.8%, AP$1\times31$7 by 4.7%, and AP$1\times31$8 by 5.2% (Zhang et al., 3 Apr 2026). The same table separates the subcomponents: WPM alone improves AP to 29.4%, LDConv + WPM reaches 31.3%, and the full model including MDDC achieves 31.6% AP (Zhang et al., 3 Apr 2026). The paper interprets this as evidence that WPM provides a large contextual gain, LDConv further boosts geometric detail handling, and the two are complementary (Zhang et al., 3 Apr 2026).

The paper also compares SFPN against several alternative necks in Table 5: PAFPN, BiFPN, AFPN, RepGFPN, and MAFPN (Zhang et al., 3 Apr 2026). SFPN is reported to achieve the best overall results among them, with AP = 31.3%, AP$1\times31$9 = 52.0%, AP$31\times1$0 = 31.8%, AP$31\times1$1 = 24.6%, AP$31\times1$2 = 40.3%, and AP$31\times1$3 = 37.1% (Zhang et al., 3 Apr 2026). The paper further emphasizes that, compared with PAFPN, SFPN improves $31\times1$4 by 6.4% while reducing parameters by 1.1M, though with a moderate FLOP increase (Zhang et al., 3 Apr 2026).

The kernel-size ablation for WPM provides an additional empirical characterization of the neck design. The tested kernel configurations are:

• $31\times1$5
• $31\times1$6
• $31\times1$7
• $31\times1$8
• $31\times1$9 (Zhang et al., 3 Apr 2026)

The best tradeoff is reported for the 31×31 design, with $1\times1$0, $1\times1$1, $1\times1$2, and FLOPs = $1\times1$3G (Zhang et al., 3 Apr 2026). The accompanying interpretation is explicit: 3×3 and 7×7 are too local, 15×15 recovers some performance, 31×31 gives a sufficiently wide field for dense small-object scenes, and 63×63 brings only a tiny gain in AP while greatly increasing computation and causing over-smoothing / boundary overflow (Zhang et al., 3 Apr 2026).

At the full-system level, experiments on VisDrone and UAVDT report that SFFNet-X achieves 36.8 AP and 20.6 AP, respectively, while the lightweight models (N/S) maintain a balance between detection accuracy and parameter efficiency (Zhang et al., 3 Apr 2026). Since SFPN is the designated neck in that system, these results situate it within a high-performing UAV detection pipeline, though the paper attributes overall performance to the combined effect of SFPN and MDDC rather than to the neck alone (Zhang et al., 3 Apr 2026).

6. Relation to the earlier synthetic SFPN

The 2022 paper "SFPN: Synthetic FPN for Object Detection" (Zhang et al., 2022) presents a different method that nonetheless addresses a related limitation of standard FPNs. Its argument is that conventional backbones such as VGG, ResNet, DenseNet, or MobileNetV2 produce feature maps at scales such as

$1\times1$4

because of pooling or stride-2 convolutions, and that this spacing is too coarse for detection (Zhang et al., 2022). The paper states that adjacent levels separated by a factor of 2 create a scale truncation problem, so that two similar objects such as $1\times1$5 and $1\times1$6 may be routed to different prediction maps (Zhang et al., 2022).

To address this, the synthetic SFPN inserts synthetic intermediate pyramid levels using the Synthetic Fusion Module (SFM) and Synthetic Fusion Block (SFB) (Zhang et al., 2022). The SFM takes three optional inputs, linearly scales them to a common spatial size, adds them pixel-by-pixel, and applies a $1\times1$7 convolution (Zhang et al., 2022). Its operation is summarized as

$1\times1$8

with the resized inputs fused after elementwise addition (Zhang et al., 2022). An SFB consists of multiple SFMs and performs fusion in two passes: first from the first batch of layers to the second batch, then from the second batch back to the first batch, which the paper describes as “merging the features centrally and then radiating the features outward” (Zhang et al., 2022).

The method defines SFPN-3, SFPN-5, and SFPN-9, along with SFPN-5-SOL and SFPN-9-SOL, where SOL = synthetic output layers (Zhang et al., 2022). The intended effect is to create a denser pyramid with scales roughly like

$1\times1$9

instead of only powers of 2 (Zhang et al., 2022). The authors report improvements across MobileNetV2, VGG16, and ResNet50 on MS-COCO, and note that the gains are more pronounced for MobileNetV2, which they interpret as support for the claim that synthetic layers help lightweight models more strongly (Zhang et al., 2022). They also state the principal limitation directly: more layers mean more parameters and lower FPS (Zhang et al., 2022).

This earlier work is not a precursor of the 2026 synergistic SFPN in the narrow architectural sense; the two papers define different modules, use different fusion operators, and target different problem decompositions. However, they share a broader research theme: both treat standard FPN spacing and fusion as insufficiently adapted to the structure of the target detection problem. A plausible implication is that the acronym collision reflects convergence toward a common concern—namely, how to preserve information across scales without losing fine localization fidelity.

7. Significance, limitations, and recurrent themes

Within the SFFNet paper, the significance of the synergistic SFPN lies in its attempt to reconcile three demands that are frequently in tension in aerial object detection: fine detail preservation, object-shape adaptability, and long-range contextual reasoning (Zhang et al., 3 Apr 2026). Rather than adding extra low-level heads, the neck uses LDConv and WPM to enhance the feature hierarchy itself (Zhang et al., 3 Apr 2026). The ablation evidence reported in the paper supports the view that these mechanisms are complementary rather than redundant, with WPM contributing context and LDConv contributing geometry-aware detail extraction (Zhang et al., 3 Apr 2026).

The principal limitations stated or implied by the paper are tied to computation. The 31×31 WPM configuration is selected as the best tradeoff precisely because larger kernels such as 63×63 increase computation substantially and induce over-smoothing / boundary overflow (Zhang et al., 3 Apr 2026). This suggests that wide-area contextual modeling is beneficial up to a point, after which the representational gain saturates relative to cost. The paper’s discussion of avoiding extra low-level heads for efficiency reasons also situates SFPN within a resource-conscious design space, especially given that SFFNet offers six detector scales N/S/M/B/L/X for different applications or resource-constrained scenarios (Zhang et al., 3 Apr 2026).

More broadly, the synergistic SFPN exemplifies a specific trend in contemporary neck design: the move away from viewing feature pyramids as purely top-down or bottom-up routing structures and toward treating them as loci for targeted operators that preserve or recover information lost under ordinary scale aggregation. The synthetic SFPN of 2022 advances that trend through denser intermediate scales (Zhang et al., 2022), whereas the synergistic SFPN of 2026 advances it through shape-adaptive low-level processing and context-enriched high-resolution enhancement (Zhang et al., 3 Apr 2026). The two methods therefore occupy different positions within the same general research trajectory of making pyramid fusion more closely matched to object geometry, scale continuity, and scene context.