PSG-RTDETR: Towards Stable Cross-Scale Feature Fusion for Small Object Detection

Highlights

What are the main findings?

• PSG-RTDETR improves UAV small object detection by introducing a P2-aware crossscale feature fusion structure.
• The proposed Softplus-Gated BiFPN encourages smoother feature aggregation and suppresses noisy or interfering shallow features.

What are the implications of the main findings?

• Stable and adaptive feature fusion can improve small object detection accuracy under a clear GPU-side computational trade-off.
• The proposed method provides an effective solution for UAV remote-sensing detection under complex backgrounds and degraded visual conditions.

Abstract

Small object detection in UAV remote sensing remains challenging due to fine-grained feature loss during downsampling and unbalanced cross-scale feature aggregation. Conventional feature pyramid fusion may suffer from hard activation truncation or competitive weight suppression when heterogeneous feature scales are fused. To address these issues, we propose PSG-RTDETR, a P2-aware Softplus-Gated RT-DETR framework tailored for UAV-based dense small-object detection. Specifically, we integrate high-resolution P2 features into a bidirectional feature pyramid to preserve critical spatial details. Furthermore, we introduce Softplus normalization for smooth weight optimization and a sample-adaptive gating mechanism to dynamically suppress noisy shallow responses. Extensive experiments on VisDrone-DET demonstrate that, compared with the RT-DETR-ResNet18 baseline under the same ablation setting, PSG-RTDETR improves mAP_50–95 and AP_small by 4.1 and 4.7 percentage points, respectively. Further validation on HIT-UAV confirms its strong generalizability across infrared modalities. These accuracy gains involve a clear computational trade-off: GFLOPs increase from 57.20 to approximately 110.23, while FPS decreases from 126.18 to 80.02 on an RTX 3090. Thus, PSG-RTDETR is better suited to accuracy-oriented off-board or near-real-time GPU analysis than to direct deployment on resource-constrained onboard UAV devices without further compression or acceleration.

1. Introduction

Small object detection in UAV-based remote sensing remains challenging due to low pixel occupancy, severe scale variation, and complex background interference [1,2]. Although Feature Pyramid Networks (FPNs) have become a standard approach for multi-scale representation learning [3], their effectiveness remains limited for extremely small objects [4,5]. Specifically, deep features often lack sufficient spatial resolution, while shallow features preserve more detailed information but contain weaker semantic representations. As a result, small-object cues can be weakened during cross-scale feature fusion, leading to degraded detection performance.

Existing UAV small-object detection methods have made progress in multi-scale feature modeling, but their performance is still constrained by the capability of cross-scale feature fusion. Classical FPN establishes the foundation of multi-scale fusion through top-down pathways and lateral connections [3]. Subsequent methods further enhance feature aggregation efficiency and detail representation through bidirectional connections, learnable weights, and improved pyramid structures [4,6,7]. Meanwhile, studies on UAV small-object detection usually introduce shallow high-resolution features, dual-path fusion, or context enhancement modules to improve the recall of fine-grained targets [8,9,10]. Recent UAV and remote-sensing small-object detectors further indicate that improving detection performance depends not only on more complex pyramid structures, but also on balancing high-resolution detail preservation, low-level texture and high-level semantic interaction, and cross-scale semantic calibration [11,12,13,14].

However, most existing methods still rely on conventional weighting and normalization strategies, which makes it difficult to simultaneously achieve smooth weight allocation, cross-level coordination, and noise suppression during heterogeneous feature fusion [6,15]. This issue becomes more evident when shallow high-resolution features are introduced. Although high-resolution features are beneficial for preserving small-object details, they may also increase computational cost and affect real-time detection efficiency [16]. More importantly, these features contain richer texture and localization information, but they also bring stronger noise responses and a larger semantic gap. Without balanced weight allocation and effective feature selection, shallow information may not be fully exploited and can introduce additional false positives and inference burden [8,9,10,17]. Meanwhile, recent methods have attempted to alleviate shallow noise, semantic mismatch, and unstable cross-scale aggregation through clean feature extraction, stable fusion, context-gated fusion, or adaptive feature aggregation [18,19,20]. Nevertheless, detail preservation, smooth weight learning, and dynamic feature selection are still often treated as relatively separate designs, and a unified lightweight fusion mechanism for RT-DETR-based UAV small-object detection remains insufficiently explored.

To address this issue, this paper focuses on the feature fusion mechanism itself rather than simply stacking additional modules or expanding the pyramid topology. We argue that effective UAV small-object detection requires a fusion design that can simultaneously preserve high-resolution details, stabilize heterogeneous cross-scale aggregation, and selectively suppress unreliable shallow responses. Based on this motivation, we redesign the BiFPN-style neck in RT-DETR from three aspects: P2-aware detail preservation, Softplus-normalized stable fusion, and sample-adaptive gated path modulation.

Accordingly, we construct a P2-aware Softplus-Gated BiFPN neck within the RT-DETR-ResNet18 framework. First, a P2 high-resolution detection path is introduced to alleviate the detail loss of UAV small objects during downsampling. Second, a Softplus-normalized BiFPN is proposed to replace hard-truncated or overly competitive fusion weights with smooth positive constraints. Third, a sample-adaptive path gate is designed to selectively receive fused features through residual interpolation, suppressing the interference of shallow noise on deep semantics.

The main contributions of this paper are summarized as follows:

(a)

We introduce a P2 high-resolution detection path to alleviate the detail loss of UAV small objects during downsampling.

(b)

We propose a Softplus-normalized BiFPN that replaces hard-truncated or overly competitive fusion weights with smooth positive constraints.

(c)

We design a sample-adaptive path gating mechanism that selectively receives fused features through residual interpolation and suppresses the interference of shallow noise on deep semantics.

2. Methodology

2.1. Overall Architecture

We adopt RT-DETR-ResNet18 as the baseline framework. The original RT-DETR-ResNet18 follows the end-to-end detection paradigm of RT-DETR, where a ResNet18 backbone is used to extract hierarchical features, followed by feature interaction and decoding stages for object prediction. However, for UAV small object detection, this structure mainly relies on middle- and high-level features for detection, while the utilization of shallow high-resolution features remains limited. This restricts the effective preservation of fine-grained spatial details. Moreover, under complex backgrounds and severe scale variations, the original cross-scale feature interaction pathway struggles to simultaneously preserve details, maintain fusion stability, and suppress interfering responses. Therefore, the focus of this work is not to locally enhance the ResNet18 backbone, but to redesign the feature fusion pathway and information flow of RT-DETR according to the requirements of small object detection.

As shown in Figure 1, the proposed PSG-RTDETR consists of four main components: a ResNet18 backbone for hierarchical feature extraction, a P2-aware Softplus-Gated BiFPN neck for smooth cross-scale fusion, a RepC3 module for feature refinement, and an RT-DETR decoder for end-to-end prediction. Compared with the baseline model, the main structural modification lies in the neck. On the one hand, shallow P2 high-resolution features are explicitly introduced to interact bidirectionally with deep semantic features. On the other hand, the original multi-scale fusion process is reformulated through Softplus normalization and gate-based path selection, improving the balance and adaptivity of feature aggregation as well as the sensitivity to small objects. In addition, the RepC3 module is introduced to further enhance the representation capability of fused features, while the decoder retains the query-based end-to-end prediction paradigm of RT-DETR. Overall, PSG-RTDETR preserves the basic backbone and detection paradigm of RT-DETR-ResNet18, but performs a structure-level redesign of the feature fusion mechanism and information flow for small object detection.

2.2. P2-Aware Detection Framework

RT-DETR mainly relies on middle- and high-level features for detection during feature modeling, with the minimum output stride starting from stride 8, corresponding to the $C_3$–$C_5$ feature levels. For extremely small objects in UAV scenarios, such downsampling may cause irreversible loss of fine-grained geometric information at early stages, thereby weakening the localization and recognition capability of the detector [1]. To alleviate this issue, we construct an explicit P2-aware fine-grained detection framework by incorporating shallow high-resolution features into the detection pipeline. The backbone produces four hierarchical feature maps:

$$\begin{array}{ccccc}& & \hfill \{C_2,C_3,C_4,C_5\}& & \hfill \end{array}$$

(1)

Before entering SG-BiFPN, these features are projected into a unified channel space:

$$\begin{array}{ccccc}& & \hfill P_l={\varphi }_l\left(C_l\right),l\in \{2,3,4,5\}& & \hfill \end{array}$$

(2)

where ${\varphi }_l(·)$ denotes the channel-alignment convolution for the l-th feature level. In particular, P2 preserves the stride-4 high-resolution response and provides finer spatial cues for UAV small objects. As illustrated in Figure 2, the P2-aware layer establishes orthogonal complementarity between high-resolution C2 features and deep P3 semantic features. Unlike conventional structures that mainly rely on deep semantic features for implicit inference, this design explicitly introduces shallow detailed information into the detection process, providing more direct texture and localization cues for extremely small objects.

The introduction of shallow P2 features is not merely the addition of a high-resolution branch, as it also brings new challenges. Although P2 contains richer detailed information, it has a lower signal-to-noise ratio and a clear semantic gap with deeper features [5,21]. If conventional BiFPN fusion is directly applied, shallow background textures and noise responses may be amplified during cross-scale aggregation, interfering with deep semantic information and degrading the overall fusion quality. Therefore, the key issue is not simply whether to introduce P2, but how to use it properly and how to achieve stable and selective feature interaction between shallow details and deep semantics. Based on this consideration, we further design SG-BiFPN to address the fusion challenges introduced by P2 through more stable weight allocation and more effective suppression of interfering responses.

2.3. SG-BiFPN

In the original RT-DETR architecture, the neck network adopts a Transformer-based Hybrid Encoder (AIFI) to model interactions among multi-scale features [22]. This structure is effective for global semantic modeling, but its computational complexity grows quadratically with the feature sequence length. When high-resolution features such as P2 are introduced, it may lead to considerable computational and memory overhead, limiting real-time applicability. More importantly, feature interaction mechanisms centered on global attention are mainly designed for high-level semantic modeling and long-range dependency capture. They do not explicitly optimize the preservation of fine-grained spatial details or the fusion of high-resolution features for small object detection, while recent remote-sensing small object detectors often introduce task-specific lightweight spatial–channel fusion or detail-enhancement mechanisms [23].

Based on this consideration, we replace the original Hybrid Encoder with the topologically efficient BiFPN to construct a lightweight cross-scale feature fusion pathway. BiFPN enables repeated interactions among multi-level features through top-down and bottom-up bidirectional connections, providing effective cross-scale aggregation with relatively low computational cost. As shown in Figure 3, various feature pyramid network designs have been proposed to enhance multi-scale representation learning. However, after explicitly introducing high-resolution P2 features, cross-scale fusion is no longer limited to adjacent feature levels. Instead, it extends to heterogeneous features with a larger scale span, which exposes new limitations of the conventional BiFPN in this setting.

Specifically, feature fusion across a large scale span faces two main conflicts. First, features from different levels show clear differences in semantic representation and statistical distribution. Shallow features usually contain high-frequency texture patterns, whereas deeper features are more abstract and semantically stronger. This heterogeneity makes it difficult for a unified linear fusion strategy to adapt effectively [24,25]. Second, shallow features often contain background textures and interfering responses. Without an explicit selection mechanism, the weighted summation in conventional BiFPN may accumulate interfering responses during cross-scale aggregation, thereby weakening the discriminative capability of fused features [17,26].

Existing studies have improved P2 or high-resolution features, fine-grained information preservation, and feature selection from different perspectives. However, these capabilities are often designed in a decoupled manner, rather than being unified into a fusion framework that jointly emphasizes smooth weight allocation and selective aggregation. Therefore, in P2-aware cross-scale fusion, the key issue is no longer to design a more complex topology, but to construct a fusion mechanism that encourages balanced weight learning while preserving feature selection capability. Motivated by this, we propose SG-BiFPN based on the BiFPN structure, providing a unified basis for smooth weighting and gate-based path selection [27].

2.3.1. Softplus Normalized Fusion Weighting

During cross-scale feature fusion, the substantial differences in the statistical distributions of features from different levels make the balanced learning of fusion weights a critical issue [24,25]. Shallow features usually preserve richer high-resolution textures and spatial details, but their semantic representations are relatively weak. In contrast, deeper features tend to encode more abstract semantic information. This distributional heterogeneity may make conventional weighting mechanisms sensitive to activation truncation or branch competition, thereby affecting the effective aggregation of multi-scale features. In particular, if a weighting function strongly favors one branch or prematurely suppresses another, the fused representation may either amplify shallow background noise or lose useful fine-grained responses.

Existing methods typically constrain fusion weights using ReLU-L1 normalization or Softmax. However, these activation functions have clear limitations in this setting. ReLU-L1 normalization, as used in EfficientDet-style fast normalized fusion, hard-truncates negative fusion parameters before L1 normalization, so inactive branches may be slow to recover once their weights become zero. In contrast, the exponential normalization of Softmax introduces strong competition among branches, which tends to produce extreme weight distributions and suppress the contribution of some feature paths [28]. When the feature distributions differ substantially, such weighting strategies may make balanced fusion difficult [6,15].

To address this issue, we use the Softplus function to model normalized fusion weights, as shown in Figure 4. For each SG-BiFPN fusion node, the two aligned input features are denoted as $X=\{X_0,X_1\}$. The learnable scalar weight of the i-th branch is denoted as ${\alpha }_i$. Softplus first maps it to a positive weight:

$$\begin{array}{ccccc}& & \hfill {\widehat{w}}_i=Softplus\left({\alpha }_i\right)=ln(1+e^{{\alpha }_i}),i\in \{0,1\}& & \hfill \end{array}$$

(3)

where ${\widehat{w}}_i$ is the non-negative weight obtained after Softplus activation. The normalized fusion weight and fused candidate feature are then computed as follows:

$$\begin{array}{ccccc}& & \hfill w_i={\displaystyle \frac{{\widehat{w}}_i}{{\sum }_{j=0}^1{\widehat{w}}_j+\epsilon }},F={\displaystyle \sum _{i=0}^1}{w}_i{X}_i=w_0{X}_0+w_1{X}_1& & \hfill \end{array}$$

(4)

where $X_i$ denotes the i-th input feature, F is the Softplus-normalized fused feature, and $\epsilon$ is a numerical stability term.

The equations above show that Softplus first maps each learnable fusion parameter to a positive response and then normalizes the responses across branches. Because Softplus is continuously differentiable over its entire domain, it avoids the hard truncation of negative responses caused by ReLU. Its monotonic and strictly positive properties allow each branch to retain a non-zero response, reducing the risk that a feature path is prematurely deactivated by the weighting function itself. Compared with the strong competition introduced by Softmax, Softplus normalization tends to produce smoother and more balanced weight distributions.

Since Softplus becomes approximately linear for large positive inputs, its advantage over ReLU-L1 is not assumed theoretically. Instead, we evaluate it empirically under the same P2-enhanced BiFPN setting in Section 3.2.3, including the Softplus-based fusion-weight distribution and smoothness analysis and the quantitative comparison of Softplus, Softmax, and ReLU-L1 normalization. In this formulation, Softplus mainly controls the smoothness and positivity of fusion-weight allocation, while feature selectivity is further complemented by the subsequent gate.

2.3.2. Gate-Based Path Selection

In multi-scale feature fusion, conventional weighted summation or normalized weighting mechanisms are essentially passive aggregation strategies, where all input branches are forced to participate in the fusion process. Although such strategies can integrate information from multiple levels to some extent, they lack the ability to evaluate feature effectiveness. As a result, irrelevant or interfering responses may be introduced during aggregation, reducing the discriminability of the fused features. Therefore, relying only on weight allocation is insufficient to address information interference in cross-scale fusion [29,30].

To address this issue, we introduce a lightweight gate-based path selection mechanism at each fusion node. Unlike conventional weighted summation, the proposed gate does not simply rescale each input branch independently. Instead, it controls how much of the fused candidate feature should be accepted by the main path. The gate is not a hand-crafted foreground/background classifier and does not explicitly label a shallow response as either a small-object feature or background noise. Rather, its parameters are optimized implicitly through the detection loss, so that the learned coefficient reflects whether the fused candidate is beneficial for the final detection objective under the current sample context.

The gate is generated from the main-path feature $X_0$. First, global average pooling is applied to obtain a compact context descriptor:

$$\begin{array}{ccccc}& & \hfill z=GAP\left(X_0\right)& & \hfill \end{array}$$

(5)

Similar to squeeze-excitation style attention, the descriptor is then passed through two lightweight $1\times 1$ convolutional layers and a Sigmoid activation. Different from channel reweighting, however, this gate is used to estimate the residual update strength from the Softplus-fused candidate to the main path:

$$\begin{array}{ccccc}& & \hfill \begin{array}{cc}\hfill h& =ReLU\left({Conv}_{1\times 1}^{C\to C_g}\left(z\right)\right),\hfill \\ \hfill s& ={Conv}_{1\times 1}^{C_g\to 1}\left(h\right),\hfill \\ \hfill g& =\sigma \left(s\right).\hfill \end{array}& & \hfill \end{array}$$

(6)

Equivalently, the gate generation process can be written compactly as:

$$g=\sigma ({Conv}_{1\times 1}^{C_g\to 1}(ReLU({Conv}_{1\times 1}^{C\to C_g}(GAP(X_0)))))$$

(7)

The final output is obtained by residual interpolation:

$$\begin{array}{ccccc}& & \hfill Y=X_0+g\odot (F-X_0)& & \hfill \end{array}$$

(8)

where $X_0$ and F denote the main-path feature and the Softplus-fused candidate, respectively, and Y is the gated output. s is the scalar gate logit produced by the second $1\times 1$ convolution, $g=\sigma \left(s\right)$ is the resulting gating coefficient, $C_g$ denotes the hidden gate dimension, $\sigma (·)$ is Sigmoid, the bias terms are absorbed into the convolutional transformations, and ⊙ denotes broadcast element-wise multiplication.

The gate controls the update strength of the Softplus-fused feature to the main path. This process neither regenerates a new feature representation nor simply rescales each branch independently; instead, it performs residual interpolation between the main-path feature and the fused candidate. As illustrated in Figure 5, the gate-based path selection mechanism enables sample-adaptive control of the fusion process. When the fused information is unreliable or dominated by background interference, a smaller gate value makes the output closer to $X_0$, thereby reducing noise propagation. When the fused information contains useful fine-grained cues, a larger gate value allows more cross-scale information to be incorporated. The learned behavior of this mechanism is further examined in Section 3.2.3 through the dynamic behavior of learned gate coefficients, including mean gate curves, coefficient distributions, and coefficient-change heatmaps.

2.3.3. Synergistic Mechanism of SG-BiFPN

The above Softplus normalization and path gating are not two independent modules. Instead, they form a unified joint modulation mechanism during feature fusion. Specifically, Softplus continuously and smoothly reallocates the response magnitudes of different branches, while the gating unit controls how strongly the fused candidate feature should update the main path. Together, they determine the actual contribution of cross-scale information to the output feature.

From the perspective of fusion formulation, conventional methods usually integrate cross-scale information through direct weighted summation, which lacks explicit control over feature flow. In contrast, SG-BiFPN reformulates feature fusion as smooth normalized aggregation followed by controlled residual updating. In this mechanism, Softplus first constructs a fused candidate with positive normalized weights, and the gate then selectively regulates the update strength from this candidate to the main-path feature.

For clarity, consider an arbitrary SG-BiFPN fusion node with N aligned inputs. The cooperative fusion process can be expressed as follows:

$$\begin{array}{ccccc}& & \hfill Y=X_0+g\left(X_0\right)\odot \left({\displaystyle \sum _{i=0}^{N-1}}{w}_i{X}_i-X_0\right)& & \hfill \end{array}$$

(9)

where $X_0$ is the main-path feature, $w_i$ is the Softplus-normalized fusion weight defined above, and $g\left(X_0\right)$ is the sample-adaptive path gate generated from $X_0$. In this work, each SG-BiFPN node uses two aligned inputs, and thus the practical form is as follows:

$$\begin{array}{ccccc}& & \hfill Y=X_0+g\left(X_0\right)\odot (w_0{X}_0+w_1{X}_1-X_0)& & \hfill \end{array}$$

(10)

Equivalently, the node output can be interpreted as follows:

$$\begin{array}{ccccc}& & \hfill Y=(1-g)\odot X_0+g\odot (w_0{X}_0+w_1{X}_1)& & \hfill \end{array}$$

(11)

Here, g is an abbreviated form of $g\left(X_0\right)$. In other words, Softplus determines how multi-scale features are smoothly fused, whereas the gate determines how much of the fused result should be accepted. As illustrated in Figure 6, SG-BiFPN first propagates deep semantic information in a top-down manner and then returns high-resolution detail information through a bottom-up path.

The output feature set passed to the decoder is as follows:

$$\begin{array}{ccccc}& & \hfill {\mathcal{P}}^{out}=\{\mathrm{P}{2}^{out},P_3^{out},P_4^{out},P_5^{out}\}& & \hfill \end{array}$$

(12)

Based on this design, the fusion process can maintain smooth weight allocation while dynamically suppressing ineffective fused responses. As a result, a more stable and controllable information flow is formed during cross-scale feature aggregation. The synergy between stability and selectivity provides a more consistent and high-quality feature representation basis for the subsequent decoding stage.

2.4. Decoder

The output feature set ${\mathcal{P}}^{out}$ refined by SG-BiFPN, namely $\mathrm{P}{2}^{out}$, $P_3^{out}$, $P_4^{out}$, and $P_5^{out}$, is fed into the decoder in a unified manner. Unlike the original RT-DETR, which directly uses less aligned multi-scale features, the proposed feature representations achieve more consistent cross-scale alignment between spatial details and semantic information. This provides a higher-quality input basis for subsequent object queries.

The decoder still follows the dynamic query mechanism of RT-DETR. However, benefiting from the joint optimization of the P2-aware architecture and SG-BiFPN, the initialized Object Queries no longer interact with coarse or semantically inconsistent feature maps. Instead, they operate in a feature space with higher information density and more controlled interference. This enables the query vectors to more accurately capture the spatial positions and structural information of small objects during multi-head self-attention and cross-attention.

Based on this improved feature representation, the IoU-aware query matching strategy can facilitate faster convergence from coarse candidates to fine-grained regression. Meanwhile, the sparse query update mechanism, i.e., Top-k selection, helps control the growth of computational complexity while maintaining high recall and improving the response to fine-grained targets [22].

3. Experiments

This section introduces the datasets, evaluation metrics, and parameter settings used in the experiments. By combining quantitative and qualitative analyses, we compare the proposed method with other state-of-the-art detection approaches.

3.1. Datasets

The VisDrone2019-DET dataset is the object detection subset of the VisDrone2019 benchmark, designed for UAV-view object detection in complex real-world scenes [2]. The dataset contains aerial images collected from urban roads, intersections, residential areas, and other public environments, where objects often suffer from dense distribution, large-scale variation, occlusion, and cluttered backgrounds. In this work, the processed dataset contains 6471 training images, 548 validation images, and 1610 test images. As shown in Figure 7, the dataset provides bounding-box annotations for 10 categories, namely pedestrian, people, bicycle, car, van, truck, tricycle, awning-tricycle, bus, and motor. The label statistics in Figure 7 are computed from the processed VisDrone2019-DET annotations [2]; they present the spatial distribution and normalized box-size correlogram of annotated objects, showing that a large proportion of targets occupy only a small image region. These characteristics are consistent with the small-object, dense-scene detection challenges addressed in this work.

The HIT-UAV dataset is an infrared thermal UAV object detection dataset collected from high-altitude aerial perspectives [31]. It contains thermal images captured in diverse scenes such as roads, schools, parking lots, and playgrounds, and is characterized by small targets, low-resolution object appearances, and complex aerial backgrounds. In this work, the processed dataset contains 2029 training images and 869 validation images. The annotated target categories include Person, Car, Bicycle, and OtherVehicle, while DontCare regions are ignored during training and evaluation.

3.2. Experimental Results on VisDrone Dataset

This section systematically evaluates the proposed method on the VisDrone2019 dataset. First, we compare the proposed method with the RT-DETR-ResNet18 baseline and other mainstream detection methods. Then, ablation studies are conducted to analyze the incremental contributions of BiFPN, the P2 feature layer, and the SG-BiFPN module. Finally, training curves and qualitative visualization results are used to further analyze the strengths and limitations of the proposed method in UAV scenarios.

The proposed framework is implemented based on Ultralytics RT-DETR (Ultralytics version 8.3.169), using an ImageNet-pretrained ResNet18 as the backbone. On this basis, the P2-aware branch and the Softplus-Gated BiFPN neck are introduced [22]. All experiments are conducted for 220 epochs on a single NVIDIA GeForce RTX 3090 GPU (NVIDIA Corporation, Santa Clara, CA, USA). AdamW is used as the optimizer, and the input image resolution is set to $640\times 640$ [32]. To better preserve small-object details and avoid synthetic context changes caused by heavy image composition, Mosaic and MixUp augmentations are not used during training. This setting is applied consistently to the baseline, all ablation models, and all dataset experiments, so the comparison does not favor the proposed modules alone. Instead, HSV jittering, scaling, and random erasing are adopted as the main augmentation strategies. Since this work does not isolate augmentation policies as an independent variable, the influence of Mosaic and MixUp on UAV small-object detection will be further studied in future work.

For fair comparison, all baseline models are evaluated under the same dataset split, input resolution, and evaluation protocol. The reported YOLO and RT-DETR baseline results are reproduced using official or publicly available implementations and configurations rather than directly copied from the original papers. Since some recent detectors, such as YOLO11, YOLO12, and RT-DETR++, are available mainly as arXiv preprints or official software releases, their citations are retained for traceability, while the numerical comparisons are based on reproducible model implementations.

Model performance is comprehensively evaluated using standard COCO metrics. Specifically, mAP_50–95 is used to measure overall localization and detection accuracy, while AP_small is used to evaluate small object detection capability. In addition, GFLOPs and FPS are reported to measure computational complexity and inference speed [33]. mAP_50–95 denotes mean average precision averaged over IoU thresholds from 0.50 to 0.95; mAP₅₀ and mAP₇₅ denote AP at IoU thresholds of 0.50 and 0.75, respectively; AP_s, AP_m, and AP_l denote AP for small, medium, and large objects, respectively. Following the COCO evaluation protocol, small objects are defined as instances with a bounding-box area smaller than 32² pixels, medium objects as those with an area from 32² to 96² pixels, and large objects as those with an area larger than 96² pixels. Under the 640 × 640 input resolution used in this work, the COCO small-object threshold corresponds to 1024 pixels, or approximately 0.25% of the image area. P and R denote precision and recall; GFLOPs denotes giga floating-point operations; FPS denotes frames per second. All mAP/AP/P/R values are reported as decimals.

3.2.1. Overall Performance Comparison

As shown in

Table 1. Comparison of different detection models.

Model	mAP50–95	mAP50	mAP75	APs	APm	APl	P	R	GFLOPs	FPS
YOLOv10-L [34]	0.22	0.37	0.22	0.13	0.34	0.41	0.50	0.39	120.05	198.45
YOLOv12-L [35]	0.26	0.43	0.27	0.16	0.40	0.47	0.57	0.43	88.58	156.57
YOLOv9-C [36]	0.27	0.44	0.27	0.16	0.41	0.54	0.56	0.44	102.36	235.79
YOLO11-L [37]	0.27	0.44	0.28	0.16	0.41	0.53	0.57	0.44	86.62	234.86
RT-DETR-X [22]	0.27	0.46	0.28	0.19	0.38	0.40	0.63	0.47	222.51	62.19
RT-DETR-ResNet18 [22]	0.26	0.43	0.26	0.18	0.36	0.40	0.60	0.46	57.20	126.18
RT-DETR-ResNet50 [22]	0.27	0.44	0.27	0.19	0.37	0.38	0.62	0.46	125.66	103.17
PSG-RTDETR	0.30	0.48	0.31	0.23	0.40	0.40	0.63	0.51	110.23	80.02

, we compare PSG-RTDETR with mainstream detectors, including the YOLO series (YOLOv9–v12) and RT-DETR variants. The comparison with YOLO-series detectors is intended as a reproducible experimental baseline comparison. For recent models whose primary descriptions are arXiv preprints or official software releases, we retain the corresponding citations but rely on official implementations and consistent evaluation settings for the reported values.

Compared with the RT-DETR-ResNet18 baseline, PSG-RTDETR improves mAP_50–95 from 0.26 to 0.30 and AP_small from 0.18 to 0.23, confirming the benefit of high-resolution feature preservation and stable cross-scale fusion. However, this improvement introduces a clear computational cost. GFLOPs increase from 57.20 to 110.23, and FPS decreases from 126.18 to 80.02 on the same hardware. Therefore, PSG-RTDETR should be interpreted as an accuracy-oriented design for UAV small-object detection rather than a purely lightweight detector.

Compared with YOLO-based detectors, PSG-RTDETR shows a clear advantage in small object detection. Although YOLO models achieve higher inference speed, their AP_small is generally around 0.16, suggesting limited capability in capturing fine-grained features in UAV scenarios. In contrast, PSG-RTDETR reaches 0.23 in AP_small, indicating that the integration of high-resolution features and cross-scale interaction is important for dense small object detection.

In addition, compared with larger RT-DETR variants such as RT-DETR-ResNet50 and RT-DETR-X, PSG-RTDETR achieves higher accuracy with lower computational cost than RT-DETR-X but remains more expensive than RT-DETR-ResNet18. This suggests that, for tasks dominated by small objects, optimizing the feature fusion mechanism can be more effective than simply increasing model capacity, while its deployment cost still needs to be considered.

3.2.2. Ablation Study

To further understand the role of each component, we analyze the ablation results from a problem–solution perspective.

First, introducing the P2 branch brings the most notable performance improvement. The mAP_50–95 increases from 0.2584 to 0.2826, corresponding to a gain of 2.42 percentage points, and AP_small increases from 0.1796 to 0.2126, corresponding to a gain of 3.30 percentage points. This indicates that the main bottleneck of the baseline lies in insufficient spatial resolution, where small objects can be easily lost during downsampling. By incorporating high-resolution P2 features, this resolution bottleneck is alleviated, leading to improved small object recall.

Based on the P2-enhanced structure, Softplus normalization further improves mAP_50–95 by 0.80 percentage points and mAP₇₅ by 1.26 percentage points. Since this incremental gain is relatively small, it should be interpreted as supportive evidence under the same training and evaluation setting rather than as a definitive estimate of run-to-run robustness. The result suggests that, after high-resolution features are introduced, hard-truncated or highly competitive weight allocation during multi-scale fusion becomes another limiting factor. Softplus provides smooth and strictly positive weight distributions, which helps maintain balanced branch participation and improves fusion-weight allocation.

Finally, introducing the gating mechanism further improves mAP_50–95 by 0.90 percentage points and AP_small by 1.38 percentage points, while maintaining comparable inference efficiency. Although the mAP_50–95 gain is also modest, the improvement in AP_small indicates that noise and redundant information contained in shallow high-resolution features may degrade performance if not properly controlled. The gating mechanism performs selective feature filtering by reducing the influence of noise-dominated or low-contribution responses, thereby improving detection accuracy with limited additional overhead.

From the perspective of computational overhead,

Table 2. Ablation study on baseline components.

Model	mAP50–95	mAP50	mAP75	APs	APm	APl	P	R	GFLOPs	FPS
RT-DETR-ResNet18	0.2584	0.4312	0.2601	0.1796	0.3600	0.4025	0.6013	0.4553	57.204	126.18
+BiFPN	0.2552	0.4281	0.2568	0.1846	0.3413	0.3762	0.6027	0.4542	49.312	134.18
+P2	0.2826	0.4710	0.2870	0.2126	0.3732	0.4011	0.6307	0.4937	110.210	83.67
+Softplus	0.2906	0.4755	0.2996	0.2132	0.3904	0.4009	0.6270	0.4970	110.210	81.86
+Gate	0.2996	0.4842	0.3105	0.2270	0.3955	0.4045	0.6316	0.5058	110.232	80.02

shows that replacing the original neck with BiFPN reduces GFLOPs from 57.204 to 49.312. The main additional cost is introduced by the P2 branch, where GFLOPs increase from 49.312 to 110.210, corresponding to an increase of 60.898 GFLOPs. In contrast, Softplus normalization introduces almost no additional floating-point computation, as GFLOPs remain 110.210 after replacing the fusion-weight normalization strategy. The gate module adds only a small overhead, increasing GFLOPs from 110.210 to 110.232, i.e., 0.022 GFLOPs. Overall, the full model increases GFLOPs by approximately 53.028 compared with RT-DETR-ResNet18, indicating that the P2 high-resolution branch is the dominant source of additional computation.

Overall, the final +Gate model improves mAP_50–95 from 0.2584 to 0.2996 and AP_small from 0.1796 to 0.2270, corresponding to gains of 4.12 and 4.74 percentage points, respectively. These overall gains are more substantial than the individual sub-one-point ablation steps, but the latter should still be interpreted cautiously because multi-seed variance is not reported. The improvements do not simply come from module stacking. Instead, they result from systematically addressing three key issues: resolution deficiency, unbalanced fusion weighting, and interference from noisy or redundant feature paths.

3.2.3. Visualization Analysis

Figure 8 compares the convergence behavior of different model variants. As shown in Figure 8a, all models improve rapidly in the early training stage and gradually become stable in the later stage. The RT-DETR-ResNet18 baseline and plain BiFPN form the lower performance range, while adding the P2 branch brings a clear improvement. On this basis, Softplus normalization further increases mAP_50–95, and the full model with both Softplus and Gate achieves the highest final accuracy. The loss curves in Figure 8b show a consistent trend: the full model reaches the lowest training loss, indicating that the proposed components improve not only detection accuracy but also optimization quality.

Figure 9 and

Table 3. Quantitative comparison of normalization strategies with the P2-enhanced BiFPN.

Normalization	mAP50–95	mAP50	mAP75	APs	APm	APl	P	R	GFLOPs	FPS
Softplus	0.2906	0.4755	0.2995	0.2132	0.3903	0.4009	0.6271	0.4970	110.210	78.34
Softmax	0.2862	0.4676	0.2958	0.2121	0.3806	0.4258	0.6179	0.4936	110.210	82.25
ReLU-L1	0.2837	0.4651	0.2891	0.2144	0.3750	0.4407	0.6191	0.4860	110.210	79.32

further explain why Softplus improves multi-scale feature fusion. In the normalization comparison, Softplus, Softmax, and ReLU-L1 are evaluated under the same P2-enhanced BiFPN architecture, and only the fusion-weight normalization strategy is changed. Here, ReLU-L1 denotes the EfficientDet-style fast normalized fusion strategy, where ReLU is used to enforce non-negative weights before L1 normalization. Softplus obtains a mAP_50–95 of 0.2906, outperforming Softmax (0.2862) and ReLU-L1 normalization (0.2837). It also achieves the highest mAP₅₀, mAP₇₅, AP_m, precision, and recall among the compared normalization strategies. Since the three strategies have the same GFLOPs, the improvement mainly comes from the normalization design rather than additional computational complexity. Figure 9a shows that Softplus produces a more moderate distribution of normalized fusion weights, reducing the risk that one feature path becomes overly dominant. Figure 9b shows that the dispersion of the fusion weights remains consistently lower than that of BiFPN+P2 during training. This result should be interpreted as evidence of smoother and more balanced weight allocation, rather than as a standalone proof of numerical training stability or gradient robustness. The training dynamics in Figure 8, especially the loss curve in Figure 8b, provide complementary evidence of convergence behavior.

Figure 10 visualizes the dynamic selection behavior learned by the gating mechanism. The mean gate curves in Figure 10a show that different fusion paths do not share a uniform activation pattern. Some paths remain close to full activation throughout training, indicating stable and consistently useful information flow, whereas other paths gradually increase from low or moderate values, suggesting that their contribution becomes more selective as training progresses. The delta heatmap in Figure 10b further shows that the gate coefficients of different paths evolve in different directions relative to the early stage. The distribution in Figure 10c confirms this non-uniform behavior: several gates have narrow distributions near high activation values, while others maintain wider or lower-valued distributions. These observations indicate that the gate module learns path-dependent residual selectivity rather than applying a fixed global scaling factor. At the same time, gates that remain consistently low or show little contribution may also reveal redundant fusion paths, which provides a useful clue for future pruning and lightweight neck redesign.

Figure 11 presents a qualitative Grad-CAM comparison under short- and long-exposure conditions. Each row corresponds to the same scene under different exposure settings. It can be observed that PSG-RTDETR maintains clear target-related activation responses in regions with dense traffic objects, such as vehicles and pedestrians, under both exposure conditions. Although the brightness distribution changes noticeably, the model responses are not dominated by background interference caused by exposure variation. Instead, the highlighted regions still mainly cover roads and traffic areas related to the detection targets. This qualitative exposure analysis is complementary to the perturbation experiments in

Table 4. Comparison under synthetic perturbations and real low-light/no-light VisDrone scenarios.

Model	Dataset	Condition	Images	P	R	mAP50	mAP50–95
RT-DETR-ResNet18	VisDrone	Original	640	0.65	0.50	0.46	0.28
	VisDrone	Gaussian Noise	640	0.66	0.49	0.46	0.28
	VisDrone	Gaussian Blur	640	0.58	0.51	0.44	0.27
	VisDrone	High Brightness	640	0.67	0.48	0.45	0.27
	VisDrone	Low Brightness	640	0.64	0.44	0.43	0.26
	VisDrone	Salt & Pepper	640	0.39	0.39	0.34	0.20
	VisDrone	Real Low-light	640	0.63	0.47	0.50	0.33
	VisDrone	Extreme Dark/No-light	640	0.62	0.46	0.49	0.31
PSG-RTDETR	VisDrone	Original	640	0.70	0.55	0.51	0.33
	VisDrone	Gaussian Noise	640	0.71	0.54	0.51	0.32
	VisDrone	Gaussian Blur	640	0.64	0.55	0.48	0.30
	VisDrone	High Brightness	640	0.71	0.53	0.50	0.31
	VisDrone	Low Brightness	640	0.69	0.50	0.47	0.30
	VisDrone	Salt & Pepper	640	0.33	0.43	0.41	0.26
	VisDrone	Real Low-light	640	0.71	0.60	0.63	0.40
	VisDrone	Extreme Dark/No-light	640	0.68	0.53	0.56	0.36

. Figure 11 illustrates feature-response behavior under real exposure variation, whereas Table 4 quantitatively reports synthetic noise, blur, brightness changes, and reallow-light/no-light scenarios. Together, they provide visual and quantitative evidence for the robustness of the proposed feature fusion mechanism in UAV scenarios [38].

3.2.4. Perturbation Scenarios

To evaluate the robustness of the proposed model, we conduct experiments under two types of degraded visual conditions. The first group consists of synthetic perturbations, including Gaussian noise, Gaussian blur, high brightness, low brightness, and Salt-and-Pepper noise. The second group uses real adverse-illumination VisDrone images, including real low-light and extreme dark/no-light scenes. As shown in Table 4, PSG-RTDETR outperforms the RT-DETR-ResNet18 baseline in terms of mAP_50–95 under different degradation conditions. On the original validation set, PSG-RTDETR achieves an mAP_50–95 of 0.33, while the baseline obtains 0.28.

Under low-brightness conditions, both models suffer from performance degradation due to reduced image contrast and weakened feature responses. Nevertheless, PSG-RTDETR maintains an mAP_50–95 of 0.30, whereas the baseline drops to 0.26. This suggests that the proposed feature fusion mechanism helps preserve more effective feature representations under weak signal conditions.

The proposed model also shows stronger robustness under noise and blur perturbations in terms of mAP. Under Gaussian noise and high brightness, PSG-RTDETR only shows slight performance degradation compared with the original setting. Under Salt-and-Pepper noise, PSG-RTDETR improves recall from 0.39 to 0.43, mAP₅₀ from 0.34 to 0.41, and mAP_50–95 from 0.20 to 0.26, while precision decreases from 0.39 to 0.33. This indicates a precision–recall trade-off under sparse impulse noise, rather than a uniform improvement across all metrics.

For real adverse-illumination images, PSG-RTDETR also maintains more stable detection performance. In the real low-light subset, PSG-RTDETR obtains an mAP_50–95 of 0.40, compared with 0.33 for RT-DETR-ResNet18. In the more challenging extreme dark/no-light subset, PSG-RTDETR still achieves an mAP_50–95 of 0.36, whereas RT-DETR-ResNet18 obtains 0.31. The proposed model also shows higher precision and recall in both real low-light and near no-light conditions, indicating that it preserves more reliable target responses when image contrast and object visibility are substantially weakened.

These results are consistent with the design motivation of SG-BiFPN. Softplus normalization provides smooth and positive fusion weights, which helps stabilize cross-scale feature aggregation under degraded inputs. Meanwhile, the gating mechanism introduces sample-adaptive residual selectivity and reduces the participation of low-quality or interfering responses in feature fusion. Therefore, the proposed model achieves higher performance under normal conditions and shows improved robustness under both synthetic perturbations and real adverse-illumination scenarios. These findings are consistent with recent studies on illumination-adaptive and adverse-condition object detection, where robust feature enhancement is important for degraded visual inputs [38,39].

Nevertheless, extreme no-light conditions remain challenging for a single-input detector because target appearance cues may become severely weakened or partially unavailable. In practical UAV perception systems, collaboration among multiple UAV cameras or complementary sensors may provide additional viewpoints, temporal cues, or modality information, which could further improve robustness under severe illumination degradation. This direction is not experimentally evaluated in this work and is left for future study.

3.3. Extended Experiments

To further evaluate the generalization capability of the proposed method, we conduct additional experiments on the HIT-UAV infrared dataset. As shown in

Table 5. Performance comparison of PSG-RTDETR with mainstream detection models (including YOLOv10/11/12 series and RT-DETR baseline) on the HIT-UAV dataset.

Model	mAP50–95	mAP50	mAP75	APsmall	APmedium	APlarge	P	R	GFLOPs	FPS
PSG-RTDETR	0.58	0.87	0.66	0.46	0.61	0.64	0.89	0.87	110.13	98.87
YOLO11-N	0.52	0.83	0.58	0.37	0.56	0.61	0.82	0.78	6.32	1004.03
YOLO12-N	0.53	0.84	0.57	0.37	0.57	0.61	0.86	0.79	6.32	656.43
YOLOv10-N	0.49	0.78	0.54	0.35	0.53	0.60	0.79	0.73	6.53	1002.73
RT-DETR-ResNet50	0.56	0.85	0.64	0.43	0.59	0.60	0.90	0.84	125.64	103.17
RT-DETR-ResNet18	0.57	0.86	0.64	0.44	0.59	0.63	0.90	0.84	57.18	183.71

, PSG-RTDETR achieves the best overall detection accuracy on the HIT-UAV dataset, with an mAP_50–95 of 0.58 and an mAP₅₀ of 0.87. Compared with lightweight YOLO detectors, PSG-RTDETR shows clear advantages in detection accuracy. In particular, it achieves an AP_small of 0.46, whereas YOLO11-N and YOLO12-N both obtain 0.37. This indicates that the proposed feature fusion mechanism remains effective in infrared small object detection, where targets usually have weaker textures and lower contrast [31]. This also agrees with recent remote-sensing small object detection studies that emphasize fine-grained feature enhancement and lightweight fusion [23].

Compared with RT-DETR-ResNet18 and RT-DETR-ResNet50, PSG-RTDETR also achieves higher mAP_50–95 and AP_small. The improvement in AP_small suggests that introducing high-resolution P2 features and stable cross-scale fusion is beneficial not only for visible-light UAV images, but also for infrared aerial scenarios. This further supports the generalization capability of the proposed structure across different sensing modalities.

In terms of recall, PSG-RTDETR reaches 0.87, outperforming the compared YOLO models and matching or slightly exceeding the RT-DETR variants. Since missed detections are particularly critical in infrared security monitoring and UAV reconnaissance applications, the higher recall further indicates the practical value of the proposed method.

Although PSG-RTDETR runs at 98.87 FPS on the RTX 3090, it is substantially slower than lightweight YOLO models on the HIT-UAV dataset. In addition, compared with RT-DETR-ResNet50, PSG-RTDETR achieves higher detection accuracy with lower computational cost. These results indicate that the proposed method favors detection accuracy and recall over maximum inference speed on infrared UAV data.

4. Discussion

The performance gains of PSG-RTDETR demonstrate the necessity of explicit fine-grained feature modeling in UAV-based small object detection. Mechanistically, incorporating P2 features restores crucial spatial details lost during conventional downsampling. However, shallow high-resolution features inherently bring background noise and a wide semantic gap when fused with deep layers. Our Softplus normalization encourages smoother and more balanced fusion-weight allocation, while the adaptive gating mechanism dynamically suppresses non-informative or noisy paths during cross-scale aggregation.

Despite these advantages, introducing P2 features and bidirectional connections inevitably scales up spatial feature dimensions, leading to a substantial increase in memory bandwidth usage and computational complexity. In the VisDrone experiments, GFLOPs increase from 57.20 to approximately 110.23, and FPS decreases from 126.18 to 80.02 on the RTX 3090. Consequently, while the current inference speed suffices for off-board or near-real-time GPU analysis, directly deploying the uncompressed architecture onto resource-constrained UAV onboard edge devices, such as Jetson-class platforms, remains challenging.

Furthermore, the gating mechanism is trained implicitly from data. Under extreme domain shifts—such as low-light nocturnal environments or severe background clutter—the learned residual path selection may lose reliability due to high-frequency noise. Addressing these trade-offs through lightweight neck architecture design, structured model compression, and explicit scene-aware gating strategies forms the main focus of our future work.

Another limitation concerns experimental variance. All reported results are obtained from a unified single training and evaluation run for each detector or ablation variant. We made our best effort to conduct fair comparisons under the same dataset split, input resolution, hardware platform, optimizer, and training schedule. However, because repeatedly training all mainstream detectors and ablation variants on both VisDrone-DET and HIT-UAV is computationally expensive, a complete multi-seed evaluation could not be finished within the revision period. Therefore, small incremental gains, especially those below one percentage point, should be interpreted with caution. Future work will report mean and standard deviation over at least three random seeds to further verify run-to-run robustness.

The current generalization evaluation is conducted on VisDrone-DET and HIT-UAV, covering visible-light and infrared UAV scenarios. The reported COCO metrics already include IoU-based evaluation through mAP_50–95, mAP₅₀, and mAP₇₅. Nevertheless, broader validation on additional public UAV or remote-sensing datasets and visualizations with explicitly marked small-target positions would further strengthen the evidence. In addition, the current latency and FPS measurements are obtained on an RTX 3090 GPU. Embedded-device latency on Jetson-class or other onboard UAV platforms has not been measured in this study and will be investigated in future hardware-aware deployment work. Future multi-UAV camera or multi-sensor collaboration may also help address no-light and severely degraded scenarios where single-view visual evidence is insufficient.

5. Conclusions

In this paper, we presented PSG-RTDETR, an enhanced object detection framework designed for UAV small-object recognition. By integrating high-resolution P2 features into a bidirectional feature pyramid, alongside a Softplus-Gated BiFPN, the proposed architecture improves feature detail preservation and selective cross-scale fusion. Experimental evaluations across both RGB (VisDrone-DET) and infrared (HIT-UAV) datasets validate the effectiveness and adaptability of the proposed gating and weight normalization modules.

Trade-offs exist between representation capability and computational overhead; the enhanced spatial resolution inevitably increases FLOPs and reduces inference speed. Overall, PSG-RTDETR is an accuracy-oriented design for offline and near-real-time GPU-based UAV inspection tasks where detection recall and accuracy on tiny targets are prioritized. Future efforts will target hardware-aware pruning, quantization, lightweight neck redesign, augmentation-policy isolation, explicit scene-aware gating, additional public-dataset validation, embedded-hardware latency testing, multi-UAV camera or multi-sensor collaboration, and multi-seed statistical validation to facilitate edge deployment on micro-UAV platforms and further assess result robustness.