In recent years, China drone technology has advanced rapidly, enabling wide applications in urban traffic monitoring, power line inspection, and agricultural surveillance. However, due to the high flight altitude of drones, the captured images often contain small targets with limited pixel information, leading to low detection accuracy, missed detections, and false positives. To address these challenges, I propose an improved model based on YOLOv8s, named OLF-YOLO, which is specifically tailored for China drone aerial imagery. My approach introduces three key innovations: an Omni-Dimensional Dynamic Convolution (ODConv) module to enhance feature extraction, a Large Separable Kernel Attention (LSKA) based C2f reconstruction module (L-C2f) to expand the receptive field, and a new loss function F-IoU inspired by Wise-IoU to better handle hard samples. Extensive experiments on the VisDrone2019 dataset demonstrate that OLF-YOLO achieves significant improvements over the baseline YOLOv8s, with mAP@0.50 increasing by 5.9% and mAP@0.50:0.95 by 4.0%. The model also outperforms several state-of-the-art methods while maintaining real-time inference speed, making it highly suitable for China drone-based object detection tasks.
Introduction
Drones have become indispensable tools in various fields, especially in China where drone applications are booming. Accurate object detection from drone perspectives is crucial for autonomous navigation, surveillance, and disaster response. However, the unique characteristics of drone-captured images—small object sizes, complex backgrounds, varying illumination, and occlusions—pose significant challenges. Traditional detection algorithms often suffer from low precision and high false alarm rates. Deep learning-based methods, particularly single-stage detectors like YOLO, offer a good balance between speed and accuracy, but they still struggle with small targets. To address these issues, I propose an enhanced YOLOv8s model that integrates dynamic convolutions, large kernel attention, and an adaptive loss function. This work is particularly relevant for China drone applications where real-time performance and high accuracy are simultaneously required.
The rest of this paper is organized as follows. I first review related work on drone-based object detection and recent improvements to YOLO-series models. Then I detail the proposed OLF-YOLO architecture, including the ODConv backbone enhancement, the L-C2f neck module, and the F-IoU loss function. Subsequently, I present the experimental setup, ablation studies, comparisons with mainstream methods, and generalization tests on PASCAL VOC and COCO datasets. Finally, I conclude with a summary and future directions.
Related Work
Object detection for drone imagery has attracted considerable attention. Two-stage detectors like Faster R-CNN provide high accuracy but are too slow for real-time applications. Single-stage detectors, especially the YOLO family, have become popular due to their speed. YOLOv4 introduced CSPDarknet53 and PaFPN. YOLOv5 used Mish activation and Mosaic augmentation. YOLOv8 further improved the architecture with C2f modules and anchor-free detection. However, these models were originally designed for general object detection and may underperform on drone-specific small objects.
Recent works have attempted to enhance YOLO for drone scenarios. For instance, Drone-YOLO incorporated attention mechanisms, LW-YOLOv8 reduced parameters, and TPH-YOLOv5 added transformer prediction heads. Others introduced new downsampling methods or feature fusion strategies. Despite these advances, challenges remain: small targets are difficult to extract due to limited pixel occupancy, and the loss function often treats all samples equally, ignoring the imbalance between easy and hard samples. My proposed OLF-YOLO addresses these issues from three complementary perspectives, making it a robust solution for China drone detection.
Proposed Method
Overall Architecture
The OLF-YOLO model is built upon YOLOv8s. The backbone network is enhanced by replacing several standard convolutional layers with ODConv modules, which dynamically adjust kernel weights across multiple dimensions. The neck network is redesigned by introducing a novel L-C2f module that integrates LSKA attention to enlarge the receptive field. Finally, the regression loss is replaced by the proposed F-IoU loss function. The overall network structure is illustrated in the figure below (the hyperlink image should be placed here).
Omni-Dimensional Dynamic Convolution (ODConv)
Standard convolution applies the same kernel parameters to all input samples, which is suboptimal for diverse drone images. I introduce ODConv to dynamically generate convolutional kernels conditioned on the input. ODConv applies attention mechanisms along four dimensions: spatial size, input channel, output channel, and kernel number. The output feature map is computed as:
$$
\mathbf{y} = \left( \alpha_{w1} \odot \alpha_{f1} \odot \alpha_{c1} \odot \alpha_{s1} \odot \mathbf{W}_1 + \cdots + \alpha_{wn} \odot \alpha_{fn} \odot \alpha_{cn} \odot \alpha_{sn} \odot \mathbf{W}_n \right) * \mathbf{x}
$$
where $\alpha_{si} \in \mathbb{R}^{k \times k}$, $\alpha_{wi} \in \mathbb{R}$, $\alpha_{fi} \in \mathbb{R}^{c_{in}}$, $\alpha_{ci} \in \mathbb{R}^{c_{in}}$ are attention weights for kernel $\mathbf{W}_i$ along spatial, output channel, input channel, and kernel dimensions, respectively; $\odot$ denotes element-wise multiplication; $*$ denotes convolution. By enabling multi-dimensional dynamic responses, ODConv enhances feature extraction capability for small targets in China drone imagery while slightly reducing model size.
L-C2f Module with Large Separable Kernel Attention
To capture long-range dependencies crucial for small object detection, I replace the standard C2f bottleneck in the neck with a novel L-C2f module. The core is the Large Separable Kernel Attention (LSKA) block, which decomposes a 2D depthwise-convolution kernel into cascaded 1D kernels (vertical and horizontal). This reduces computational complexity while maintaining a large receptive field. The LSKA output for a given channel is:
$$
\bar{\mathbf{Z}}^C = \sum_{H,W} \mathbf{W}_{(2d-1) \times 1}^C \left( \sum_{H,W} \mathbf{W}_{1 \times (2d-1)}^C \mathbf{F}^C \right)
$$
$$
\hat{\mathbf{Z}}^C = \sum_{H,W} \mathbf{W}_{k/d \times 1}^C \left( \sum_{H,W} \mathbf{W}_{1 \times k/d}^C \bar{\mathbf{Z}}^C \right)
$$
$$
\mathbf{A}^C = \mathbf{W}_{1 \times 1} * \hat{\mathbf{Z}}^C
$$
$$
\bar{\mathbf{F}}^C = \mathbf{A}^C \odot \mathbf{F}^C
$$
where $\mathbf{F}^C$ is the input feature of channel $C$, $\mathbf{W}_{1 \times (2d-1)}^C$ and $\mathbf{W}_{(2d-1) \times 1}^C$ are 1D depthwise convolution kernels with dilation $d$, $\mathbf{W}_{1 \times k/d}^C$ and $\mathbf{W}_{k/d \times 1}^C$ are 1D depthwise convolution kernels for compensation, and $\mathbf{W}_{1 \times 1}$ is a pointwise depthwise kernel to generate attention map $\mathbf{A}^C$. The L-C2f module uses two such LSKA blocks in place of standard convolutions, effectively enlarging the receptive field for drone-related small targets.
New Loss Function: F-IoU
YOLOv8s originally uses a combination of DFL and CIoU. CIoU is defined as:
$$
L_{CIoU} = 1 – IoU + \frac{\rho^2(b, b^{gt})}{C_w^2 + C_h^2} + \frac{4}{\pi^2} \left( \arctan\frac{w^{gt}}{h^{gt}} – \arctan\frac{w}{h} \right)^2
$$
where $IoU$ is intersection over union, $\rho$ is Euclidean distance between centers, $C_w, C_h$ are the smallest enclosing box dimensions, and $w, h$ are predicted width/height. CIoU does not differentiate between easy and hard samples, leading to suboptimal training. Inspired by Wise-IoU which dynamically assigns higher loss weights to low-IoU predictions, I propose F-IoU:
$$
L_{FIoU} = \frac{\theta}{1+\theta} R_{WIoU} L_{IoU} + \frac{1}{1+\theta} \left( \frac{(x-x^{gt})^2 + (y-y^{gt})^2}{C_w^2 + C_h^2} + \alpha v \right)
$$
$$
R_{WIoU} = \exp\left( \frac{(x-x^{gt})^2 + (y-y^{gt})^2}{W_g^2 + H_g^2} \right)
$$
where $\theta \in [0,1]$ balances two terms: the first term emphasizes IoU for hard samples (low IoU) via $R_{WIoU}$, and the second term refines geometric alignment (center distance and aspect ratio) for high-IoU predictions. The parameter $\theta$ is tuned experimentally; in my tests, $\theta=0.6$ yields the best performance on the VisDrone2019 dataset. F-IoU enables the model to focus more on challenging small targets commonly found in China drone images, improving overall detection accuracy.
Experiments
Dataset and Settings
I conduct experiments on the VisDrone2019 dataset, which consists of 8629 drone-captured images (6471 for training, 548 for validation, 1610 for testing) covering 10 object classes: pedestrian, people, bicycle, car, van, truck, tricycle, awning-tricycle, bus, and motorcycle. Small objects account for 67.2% of all instances, and about 50% are partially occluded, making it a challenging benchmark for China drone detection. All experiments are performed on an NVIDIA GeForce RTX 4080 GPU with Windows 11, PyTorch 2.0, and CUDA 11.9. Input images are resized to 640×640, batch size is 4, and training runs for 250 epochs. I use standard data augmentation including mosaic, random affine, and HSV jitter.
Evaluation Metrics
I report mean Average Precision (mAP) at IoU thresholds of 0.50 (mAP@0.50) and averaged over 0.50:0.95 with step 0.05 (mAP@0.50:0.95). Additionally, I measure model size (MB), computational cost (GFLOPs), and inference speed (FPS). Precision $P$ and recall $R$ are defined as:
$$
P = \frac{TP}{TP+FP}, \quad R = \frac{TP}{TP+FN}
$$
where $TP$, $FP$, $FN$ are true positives, false positives, and false negatives, respectively. Average precision per class is $AP = \int_0^1 P(R) dR$, and $mAP = \frac{1}{n}\sum_{k=1}^n AP(k)$.
Ablation Study of F-IoU Hyperparameter θ
I first investigate the effect of θ in F-IoU on the baseline YOLOv8s. Results are shown in Table 1.
| θ | mAP@0.50 (%) | mAP@0.50:0.95 (%) |
|---|---|---|
| 0.2 | 38.9 | 22.7 |
| 0.4 | 39.1 | 22.9 |
| 0.5 | 39.4 | 23.8 |
| 0.6 | 39.7 | 24.2 |
| 0.8 | 39.4 | 24.0 |
As seen, θ=0.6 yields the best mAP, demonstrating that a balanced emphasis between hard-sample IoU and geometric refinement is optimal. This value is used in all subsequent experiments.
Ablation Study of Each Component
To validate the effectiveness of each proposed module, I perform an ablation study starting from the baseline YOLOv8s and progressively adding ODConv, L-C2f, and F-IoU. Results are presented in Table 2.
| Method | Model Size (MB) | GFLOPs | FPS | mAP@0.50 (%) | mAP@0.50:0.95 (%) |
|---|---|---|---|---|---|
| Baseline (YOLOv8s) | 22.5 | 28.7 | 89.2 | 39.2 | 23.8 |
| +ODConv | 21.4 | 26.7 | 91.6 | 40.0 | 24.6 |
| +L-C2f | 23.1 | 30.1 | 86.3 | 41.2 | 25.1 |
| +F-IoU | 22.6 | 28.7 | 88.7 | 39.7 | 24.2 |
| +ODConv + L-C2f | 22.9 | 28.3 | 84.5 | 44.3 | 25.9 |
| +ODConv + F-IoU | 21.4 | 27.7 | 86.5 | 41.5 | 24.8 |
| +L-C2f + F-IoU | 23.0 | 30.2 | 83.8 | 43.8 | 26.5 |
| OLF-YOLO (All) | 22.0 | 28.4 | 78.4 | 45.1 | 27.8 |
Each individual component improves the baseline. ODConv reduces model size by 1.1 MB while increasing mAP@0.50 by 0.8%. L-C2f brings the largest gain of 2.0% in mAP@0.50, confirming the benefit of enlarged receptive fields. F-IoU adds a modest but consistent improvement. When all three are combined, the model achieves 45.1% mAP@0.50 and 27.8% mAP@0.50:0.95, representing a 5.9% and 4.0% increase over the baseline, respectively. Although inference speed drops to 78.4 FPS, it still exceeds the typical 50-60 FPS requirement for real-time China drone applications.
Comparison with Mainstream Detection Models
I compare OLF-YOLO with several state-of-the-art detectors on the VisDrone2019 test set, as shown in Table 3.
| Model | GFLOPs | FPS | mAP@0.50 (%) | mAP@0.50:0.95 (%) |
|---|---|---|---|---|
| YOLOv4 | 70.8 | 39.5 | 33.9 | 17.8 |
| YOLOv5s | 16.0 | 95.2 | 32.8 | 17.4 |
| YOLOv6s | 44.0 | 56.0 | 31.7 | 21.7 |
| YOLOv7-tiny | 13.3 | 69.0 | 24.3 | 12.1 |
| Cascade-RCNN | – | – | 31.9 | 16.1 |
| RetinaNet | 56.8 | – | 28.7 | 11.9 |
| Faster R-CNN | – | – | 33.1 | 16.8 |
| YOLOv8m | 78.7 | 82.4 | 43.6 | 25.5 |
| YOLOv8s | 28.7 | 89.2 | 39.2 | 23.8 |
| OLF-YOLO (Ours) | 28.4 | 78.4 | 45.1 | 27.8 |
OLF-YOLO outperforms all compared methods. Notably, it surpasses YOLOv8m (43.6% → 45.1% mAP@0.50) with only 28.4 GFLOPs compared to 78.7 GFLOPs, demonstrating superior efficiency. Compared to other drone-specific models (Table 4), my model also achieves the best results.
| Model | mAP@0.50 (%) | mAP@0.50:0.95 (%) |
|---|---|---|
| Drone-YOLO | 42.8 | 25.6 |
| LW-YOLOv8 | 43.3 | 25.5 |
| TPH-YOLOv5 | 40.9 | 23.3 |
| CRP-YOLO | 36.8 | 20.7 |
| DC-YOLOv8 | 41.5 | 23.9 |
| OLF-YOLO | 45.1 | 27.8 |
Generalization Experiments
To evaluate generalization, I test OLF-YOLO on PASCAL VOC 2012 and COCO datasets. On PASCAL VOC 2012, the model improves over YOLOv8s by 1.3% in mAP@0.50. On COCO, it achieves 0.8% gain in mAP@0.50:0.95. These results confirm that OLF-YOLO not only excels on small-target drone datasets but also generalizes well to larger objects, making it a versatile solution for China drone applications.
Conclusion
In this paper, I presented OLF-YOLO, an improved target detection model designed specifically for China drone imagery. By integrating ODConv into the backbone, L-C2f into the neck, and F-IoU as the loss function, the model effectively addresses the challenges of small target detection, including limited features, missed detections, and false positives. Experimental results on VisDrone2019 show a 5.9% and 4.0% improvement in mAP@0.50 and mAP@0.50:0.95, respectively, over the baseline YOLOv8s. The model outperforms numerous mainstream and drone-specific detectors while maintaining real-time inference speed (78.4 FPS). Ablation studies validate the contribution of each component, and generalization tests on VOC and COCO demonstrate robust performance. Future work will focus on further model lightweighting and deployment on embedded systems for practical China drone platforms.