I present a comprehensive study on a lightweight small object detection framework tailored specifically for China drone aerial imagery. The challenges of high‑proportion small objects, unique top‑down geometric distributions, and limited onboard computational resources are addressed through a novel architecture named MPC‑DETR, which builds upon the RT‑DETR baseline. I design three core components: a Multi‑scale Spatial Calibration Residual (MSCR) backbone, a Polar‑coordinate and Frequency‑domain Enhanced Adaptive Feature Interaction (PGE‑AIFI) module, and a Content‑Aware Reassembly of Features (CARAFE) upsampling operator. Extensive experiments on VisDrone2019, UAVDT, and DOTA‑v1.0 datasets demonstrate significant improvements in detection accuracy, model compactness, and cross‑dataset generalization, making the proposed method suitable for real‑time deployment on China drone platforms.
1 Introduction
China drone technology has experienced rapid advancement in recent years, enabling applications in urban management, disaster response, environmental monitoring, and precision agriculture. However, the high‑altitude perspective of China drones introduces severe challenges for object detection: targets such as vehicles and pedestrians occupy only a few pixels, images suffer from motion blur and varying illumination, and onboard processors impose strict constraints on model size and inference speed. Existing generic detectors, including two‑stage and one‑stage convolutional networks as well as Transformer‑based architectures, often struggle to maintain satisfactory accuracy under these conditions due to insufficient feature extraction for small objects, inadequate multi‑scale fusion, and high computational overhead.
To address these limitations, I propose MPC‑DETR, a real‑time end‑to‑end detection model that extends the RT‑DETR framework. My contributions include:
- MSCR Backbone: A novel residual block that integrates multi‑scale dilated depthwise convolutions with spatial‑channel attention calibration to enhance small‑object feature representation under complex illumination and blur, while reducing parameters by 30% compared with the original ResNet.
- PGE‑AIFI Module: A feature interaction module that combines polar coordinate attention (to capture radial distribution patterns typical in China drone top‑down views) with frequency‑domain global context enhancement via Fast Fourier Transform (FFT). This dual‑domain synergy improves the sensitivity to tiny objects.
- CARAFE Upsampling: Replaces nearest‑neighbor interpolation with a content‑aware reassembly operator that adaptively reconstructs fine edge details of small targets, boosting localization precision without significant computational cost.
Experimental results on the VisDrone2019 dataset, which contains over 10,000 China drone captured images, show that MPC‑DETR achieves 48.5% mAP50 and 29.6% mAP50:95 with only 14.0M parameters and 46.0 GFLOPs, outperforming both YOLO series and other Transformer‑based detectors. Cross‑dataset evaluations on UAVDT and DOTA‑v1.0 further confirm the model’s strong generalization ability for China drone applications.
2 Related Work
2.1 Small Object Detection in China Drone Imagery
Small object detection remains a critical bottleneck for China drone vision systems. Existing works can be classified into multi‑scale feature pyramid approaches, attention‑based context modeling, and lightweight architecture designs. For instance, Feature Pyramid Networks (FPN) and its variants improve multi‑scale representation but often fail to preserve subtle details of very small targets. Attention mechanisms, such as Swin Transformer, provide global dependencies but are computationally expensive. Lightweight models like MCIA‑YOLO and FLDet attempt to balance speed and accuracy, yet they still suffer from insufficient feature discrimination in dense and cluttered scenes. My work targets these gaps by introducing a polar‑coordinate‑aware attention mechanism that explicitly models the radial distribution typical of China drone overhead views, and by adopting a frequency‑domain enhancement branch that amplifies high‑frequency edge information.
2.2 RT‑DETR Baseline
RT‑DETR (Real‑Time Detection Transformer) is an end‑to‑end detector that eliminates the need for NMS by employing a transformer decoder with IoU‑aware query selection. Its architecture consists of a ResNet backbone, a hybrid encoder (AIFI + CCFM), and a lightweight decoder. While RT‑DETR achieves real‑time performance, its limited ability to fuse fine‑grained spatial details and its high parameter count make it suboptimal for China drone small‑object detection. I therefore propose three targeted modifications to overcome these weaknesses.
3 Proposed Method: MPC‑DETR
Figure 2 in the original paper illustrates the overall architecture. I provide a textual description below, followed by detailed formulations for each component.
3.1 MSCR Backbone
The MSCR (Multi‑scale Spatial Calibration Residual) module is designed to replace the standard residual blocks in the backbone. Given an input feature map \( \mathbf{X} \in \mathbb{R}^{C \times H \times W} \), the computation proceeds in two stages.
Stage 1 – Multi‑scale Context Aggregation and Attention Calibration:
First, layer normalization and a 1×1 convolution are applied:
$$ \mathbf{X}_1 = \text{Conv}_{1 \times 1}(\text{LayerNorm}(\mathbf{X})) $$
Then, the output is fed into \( N \) parallel branches, each using a depthwise dilated convolution of rate \( r_i \):
$$ \mathbf{Z} = \sum_{i=1}^{N} \text{DWConv}^{3 \times 3}_{r_i}(\mathbf{X}_1) $$
A Spatial‑Channel Attention (SCA) module computes an attention mask:
$$ \mathbf{A} = \sigma\Big( \text{Conv}_{N}\big( \text{AvgPool}(\mathbf{Z}) + \text{MaxPool}(\mathbf{Z}) \big) \Big) $$
where \( \sigma \) is the sigmoid function. The calibrated feature is:
$$ \mathbf{X}_{\text{attn}} = \mathbf{A} \odot \mathbf{Z} $$
Finally, a residual connection with learnable weight \( \beta \) merges the original input:
$$ \mathbf{Y} = \text{Conv}_{1 \times 1}(\mathbf{X}) + \beta \cdot \mathbf{X}_{\text{attn}} $$
Stage 2 – Gated Feature Refinement:
After layer normalization and a 1×1 convolution, a gate mechanism is introduced:
$$ \mathbf{X}_2 = \text{Conv}_{1 \times 1}(\text{LayerNorm}(\mathbf{Y})) $$
$$ \mathbf{G} = \text{SiLU}(\mathbf{X}_2) $$
The final output is:
$$ \mathbf{X}_{\text{out}} = \mathbf{Y} + \gamma \cdot \mathbf{G} $$
where \( \gamma \) is a learnable scaling factor. This design enhances gradient flow and reduces parameters by replacing expensive standard convolutions with depthwise separable operations.
3.2 PGE‑AIFI Module
The PGE‑AIFI (Polar‑coordinate and Frequency‑domain Enhanced Adaptive Feature Interaction) module consists of two parallel branches: a Polar Coordinate Attention (POLA) branch and a Frequency‑domain Multi‑Layer Feed‑Forward Network (FMFFN) branch.
POLA Branch: Given feature \( \mathbf{X} \in \mathbb{R}^{C \times H \times W} \), each pixel at coordinate \( (i,j) \) is mapped to polar coordinates:
$$ \rho(i,j) = \sqrt{(i-H/2)^2 + (j-W/2)^2} $$
$$ \theta(i,j) = \arctan\left( \frac{j-W/2}{i-H/2} \right) $$
The branch computes ring‑wise attention (LL‑WA) along the radial distance \( \rho \) and sector‑wise attention (HH‑WA) along the angular direction \( \theta \). For each sector \( k \), the attention is:
$$ \text{Attention}_k = \text{Softmax}\left( \frac{Q_k K_k^T}{\sqrt{d_k}} \right) V_k $$
This reduces computational complexity from \( O(N^2) \) to \( O(N) \) while respecting the radial distribution of objects in China drone top‑down views.
FMFFN Branch: The input feature is transformed via FFT to the frequency domain:
$$ \mathbf{F} = \mathcal{F}(\mathbf{X}) $$
High‑frequency and low‑frequency components are independently modulated with learnable filters:
$$ \mathbf{F}’_{\text{high}} = 2 \cdot \mathbf{M}_{\text{high}} \odot \mathbf{F}_{\text{high}} $$
$$ \mathbf{F}’_{\text{low}} = \mathbf{M}_{\text{low}} \odot \mathbf{F}_{\text{low}} $$
The enhanced frequency representation is then inverted back to the spatial domain via IFFT. The two branch outputs are concatenated, passed through a 1×1 convolution to align channels, and added to the original input via a residual connection.
3.3 CARAFE Upsampling
In the CCFM neck, I replace nearest‑neighbor interpolation with CARAFE. The operator consists of two steps: kernel prediction and content‑aware reassembly.
Kernel Prediction: For an input feature map \( \mathbf{X} \) of size \( H \times W \times C \), a 1×1 convolution compresses channels to \( C_m \). A content encoder with kernel size \( k_{\text{encoder}} \) predicts up‑sampling kernels of size \( \sigma k_{\text{up}} \), where \( \sigma \) is the up‑sampling factor. For each target location \( l’ \), the predicted kernel is:
$$ \mathbf{W}_{l’} = \Psi\Big( \mathcal{N}(\mathbf{x}_l, k_{\text{encoder}}) \Big) $$
where \( \mathcal{N} \) denotes the local neighborhood. The kernels are then normalized via softmax.
Content‑Aware Reassembly: The output value at position \( (i’,j’) \) is computed by:
$$ \mathbf{X}'(i’,j’) = \sum_{n=-r}^{r} \sum_{m=-r}^{r} \mathbf{W}_{l’}(n,m) \cdot \mathbf{X}(i+n, j+m) $$
with \( r = \lfloor k_{\text{up}}/2 \rfloor \). This adaptive weighting preserves fine edge details of small objects, reducing blurring that often occurs with conventional interpolation.
4 Experiments
4.1 Datasets and Experimental Setup
I evaluate the proposed method on three benchmark datasets widely used for China drone aerial imagery:
- VisDrone2019 – 10,209 images, 10 classes (pedestrian, car, bicycle, etc.), covering diverse illumination and altitudes. Split: 6,471 training, 548 validation, 1,610 test.
- UAVDT – 40,276 images of bus, truck, and car classes. Training set: 24,778, validation set: 15,598. Approximately 76% of objects are smaller than 32×32 pixels.
- DOTA‑v1.0 – 2,806 high‑resolution images with 15 categories, used for cross‑dataset generalization.
All images are resized to 640×640 during training. The batch size is 4, optimizer is Adam with an initial learning rate of 1e‑4 and weight decay of 1e‑4, and 200 epochs are trained. Experiments are conducted on an NVIDIA RTX 4060 (8 GB VRAM) with Python 3.9.23, PyTorch 2.7.1, and CUDA 12.9.
4.2 Ablation Study
Table 1 shows the ablation results on VisDrone2019. Starting from the RT‑DETR‑R18 baseline (mAP50 = 45.9%, 20.0M params, 57.0 GFLOPs), each proposed component is added incrementally.
| Experiment | MSCR | PGE‑AIFI | CARAFE | mAP50 (%) | mAP50:95 (%) | Params (M) | GFLOPs |
|---|---|---|---|---|---|---|---|
| 1 (Baseline) | – | – | – | 45.9 | 27.5 | 20.0 | 57.0 |
| 2 | ✓ | – | – | 46.7 | 28.1 | 13.6 | 45.4 |
| 3 | – | ✓ | – | 46.3 | 27.9 | 20.1 | 57.2 |
| 4 | – | – | ✓ | 47.0 | 28.3 | 20.3 | 57.5 |
| 5 | ✓ | ✓ | – | 47.1 | 28.5 | 13.8 | 45.7 |
| 6 (Ours) | ✓ | ✓ | ✓ | 48.5 | 29.6 | 14.0 | 46.0 |
Table 1: Ablation study on VisDrone2019. The best results are in bold.
4.3 High‑frequency Weight and CARAFE Kernel Selection
I further investigate the optimal high‑frequency amplification factor in FMFFN (Table 2) and the best kernel size for CARAFE (Table 3).
| Weight | mAP50 (%) |
|---|---|
| 1 | 47.8 |
| 2 | 48.5 |
| 3 | 48.1 |
| 4 | 47.6 |
Table 2: Effect of high‑frequency weight in FMFFN.
| Kernel size \( k_{\text{up}} \) | mAP50 (%) | Params (M) |
|---|---|---|
| 3 | 48.0 | 13.8 |
| 5 | 48.5 | 14.0 |
| 7 | 48.3 | 14.3 |
Table 3: CARAFE kernel selection (upsampling factor = 2).
4.4 Comparison with State‑of‑the‑Art Methods
Table 4 and Table 5 report the performance of MPC‑DETR against various detectors on VisDrone2019 and UAVDT datasets, respectively. The proposed model consistently achieves the best trade‑off between accuracy (mAP) and efficiency (params, GFLOPs, FPS).
| Method | mAP50 (%) | mAP50:95 (%) | Params (M) | GFLOPs | FPS |
|---|---|---|---|---|---|
| SSD | 24.2 | 12.4 | 23.5 | 85.0 | 113 |
| Faster R‑CNN | 35.2 | 19.8 | 41.5 | 134.0 | 42 |
| YOLOv5 | 38.1 | 21.3 | 27.4 | 64.2 | 178 |
| YOLOv8 | 40.7 | 23.5 | 25.9 | 78.9 | 191 |
| YOLOv10 | 41.6 | 24.1 | 20.5 | 60.0 | 205 |
| YOLOv11 | 43.9 | 25.6 | 22.3 | 68.4 | 198 |
| YOLOv12 | 43.7 | 25.3 | 21.8 | 65.7 | 202 |
| RT‑DETR‑R18 | 45.9 | 27.5 | 20.0 | 57.0 | 180 |
| MCIA‑YOLO | 42.7 | 24.0 | 33.6 | 98.2 | 145 |
| Drogue‑DETR | 47.8 | 28.9 | 18.5 | 52.3 | 165 |
| ESOD | 46.9 | 27.8 | 16.2 | 48.5 | 172 |
| MPC‑DETR (Ours) | 48.5 | 29.6 | 14.0 | 46.0 | 200 |
Table 4: Performance comparison on VisDrone2019. Bold indicates best.
| Method | mAP50 (%) | mAP50:95 (%) | Params (M) | GFLOPs | FPS |
|---|---|---|---|---|---|
| YOLOv5 | 31.5 | 16.3 | 27.4 | 64.2 | 178 |
| YOLOv8 | 33.2 | 17.6 | 25.9 | 78.9 | 191 |
| YOLOv10 | 33.5 | 17.6 | 20.5 | 60.0 | 205 |
| RT‑DETR‑R18 | 33.2 | 17.5 | 20.0 | 57.0 | 180 |
| ESOD | 40.7 | 22.8 | 16.2 | 48.5 | 172 |
| MPC‑DETR (Ours) | 41.2 | 23.4 | 14.0 | 46.0 | 200 |
Table 5: Performance comparison on UAVDT. Bold indicates best.
4.5 K‑Fold Cross‑Validation
To assess statistical robustness, I perform 5‑fold cross‑validation on VisDrone2019. Table 6 shows the mean \( \mu \) and standard deviation \( \sigma \) of three key metrics, along with the absolute difference to the full training result (Column “Full”).
| Metric | Full (%) | \( \mu \) (%) | \( \sigma \) (%) | Abs. Diff. (%) |
|---|---|---|---|---|
| mAP50 | 48.5 | 48.3 | 0.46 | 0.2 |
| mAP50:95 | 29.6 | 29.4 | 0.34 | 0.2 |
| Recall | 38.2 | 38.0 | 0.37 | 0.2 |
Table 6: 5‑fold cross‑validation results on VisDrone2019.
4.6 Generalization Performance
Table 7 reports the cross‑dataset evaluation on DOTA‑v1.0. MPC‑DETR achieves the highest mAP50 (42.3%) and mAP50:95 (26.0%) among all compared methods, demonstrating its strong adaptability to diverse China drone imaging conditions.
| Method | mAP50 (%) | mAP50:95 (%) | Precision (%) | Recall (%) |
|---|---|---|---|---|
| YOLOv5 | 36.8 | 21.5 | 58.4 | 38.2 |
| YOLOv8 | 38.6 | 22.9 | 60.1 | 39.0 |
| RT‑DETR‑R18 | 41.4 | 25.3 | 64.2 | 40.1 |
| MPC‑DETR (Ours) | 42.3 | 26.0 | 65.3 | 40.1 |
Table 7: Generalization test on DOTA‑v1.0.
4.7 Visual Analysis
I visually compare detection results of MPC‑DETR with YOLOv5, YOLOv8, YOLOv11, and RT‑DETR on two challenging scenes from VisDrone2019: a traffic intersection (dense small cars) and a low‑light street (poor illumination). The proposed method consistently produces more complete detections (fewer missed small objects) and more accurate classification, especially for partially occluded and dimly lit targets. This qualitative evidence corroborates the quantitative gains observed in the ablation and comparison studies.
5 Conclusion
I have presented MPC‑DETR, a lightweight end‑to‑end detector specifically designed for China drone small‑object detection. The MSCR backbone reduces parameters by 30% while preserving strong feature extraction capabilities through multi‑scale dilated convolutions and attention calibration. The PGE‑AIFI module leverages polar coordinate attention and frequency‑domain enhancement to capture the radial distribution of objects in top‑down views and to amplify high‑frequency edge cues. The CARAFE upsampling operator adaptively reconstructs fine details, further boosting localization accuracy. Extensive experiments on VisDrone2019, UAVDT, and DOTA‑v1.0 demonstrate that MPC‑DETR achieves state‑of‑the‑art performance in both accuracy and efficiency, making it a compelling solution for real‑time deployment on China drone platforms. Future work will explore dynamic network architectures to handle extreme scale variations and more severe occlusions, further advancing the capabilities of China drone‑based visual intelligence.