In recent years, drone technology has been widely adopted across various domains including agriculture, disaster response, and intelligent transportation. The ability to detect small objects from high-altitude aerial imagery is crucial for these applications, yet it remains a challenging task due to the dense distribution, scale variation, and complex backgrounds typical of drone-captured images. The Real-Time Detection Transformer (RT-DETR) has shown promising performance in balancing speed and accuracy, but its high computational cost and large parameter count hinder its deployment on resource-constrained drone platforms. To address these limitations, we propose a lightweight and accurate detection architecture named LA-DETR. Our approach introduces four key innovations: a Cross Hierarchical Multi-path Feature Aggregation (CHMFA) module to enrich feature representation, a Partially Convolutional Squeeze-and-Excitation (PCSE) block to reduce computational overhead while enhancing salient regions, a Spatial-to-Channel Feature Reuse (SCFR) mechanism to compensate for lost detail information in small object detection layers, and a Layer Adaptive Magnitude-based Pruning (LAMP) strategy to further eliminate redundant weights. Extensive experiments on the VisDrone2019 and HIT-UAV datasets demonstrate that LA-DETR reduces the number of parameters by 39.2% and GFLOPs by 27.5% compared to the original RT-DETR, while improving mAP50 by 0.6%. Our method achieves a favorable trade-off between detection accuracy and computational efficiency, making it well-suited for real-time drone-based object detection tasks.
Drone technology has revolutionized the way we perceive and interact with the environment. Equipped with cameras and sensors, drones can cover vast areas quickly, providing invaluable data for monitoring, surveillance, and mapping. However, due to the high altitude and wide field of view, objects in aerial images are often extremely small, sometimes occupying only a few pixels. This poses a significant challenge to conventional object detectors that rely on fixed receptive fields and downsampling operations. The RT-DETR model, while being one of the first real-time Transformer-based detectors, still suffers from excessive computational burden because of its heavy use of standard convolutions and multi-scale feature fusion. In this paper, we aim to optimize RT-DETR specifically for drone technology applications, preserving its end-to-end detection elegance while drastically reducing its resource footprint.
The remainder of this article is organized as follows. In Section 2, we detail the architecture of LA-DETR, including the CHMFA module, PCSE block, SCFR module, and the LAMP pruning strategy. Section 3 presents experimental results and analysis on two benchmark datasets. Finally, we conclude our work in Section 4.
Methodology
Overall Architecture
LA-DETR inherits the encoder-decoder paradigm of RT-DETR but introduces several modifications to the backbone, the efficient hybrid encoder, and the detection head. The backbone extracts multi-scale feature maps (S3, S4, S5). The efficient hybrid encoder comprises an attention-based intra-scale feature interaction (AIFI) module and a CNN-based cross-scale feature fusion (CCFF) module. We replace the original max-pooling layers with the CHMFA module and substitute the basic convolution blocks with PCSE blocks. Additionally, the SCFR module is integrated into the small object detection layer to recover fine-grained details. Finally, LAMP pruning is applied to compress the model further.
Cross Hierarchical Multi-path Feature Aggregation (CHMFA)
The original RT-DETR uses a max-pooling operation for downsampling, which inevitably discards critical pixel information of small objects. To mitigate this, we propose CHMFA, which aggregates features from multiple paths—including convolution, max-pooling, and average-pooling—to preserve contextual and local details simultaneously. As shown in the following figure, the input feature map \( X \in \mathbb{R}^{C \times H \times W} \) is first compressed via a 1×1 convolution to half the channels, resulting in \( X_1 \in \mathbb{R}^{C/2 \times H \times W} \). Then two parallel branches are applied.
\[
\begin{aligned}
X_{11} &= \text{MaxPool}_{3\times3}(\text{CBR}_{1\times1}(X_1)) \\
X_{12} &= \text{CBR}_{1\times1}(X_{11}) \\
X_{21} &= \text{AvgPool}_{3\times3}(\text{CBR}_{1\times1}(X_1)) \\
X_{22} &= \text{CBR}_{1\times1}(X_{21}) \\
X_3 &= \text{CBR}_{3\times3}(X_1) \\
X_{\text{out}} &= \text{CBR}_{1\times1}\big(\text{Concat}[X_1, X_{11}, X_{12}, X_{21}, X_{22}, X_3]\big)
\end{aligned}
\]
By concatenating the outputs of max-pooling (emphasizing salient features), average-pooling (smoothing context), and direct convolution (preserving local texture), the CHMFA module provides a richer representation that compensates for the loss of small-object information. Experimental results show that this design improves mAP50 by 1.1% over the baseline without significantly increasing computational cost.
| Connection Type | P (%) | R (%) | mAP50 (%) | mAP50-95 (%) |
|---|---|---|---|---|
| Parallel direct | 60.3 | 44.1 | 45.9 | 27.9 |
| Skip + parallel direct | 61.0 | 42.5 | 45.8 | 27.7 |
| Direct | 60.3 | 44.9 | 46.4 | 28.3 |
| Skip + direct (Ours) | 60.8 | 45.1 | 46.6 | 28.7 |
Partially Convolutional Squeeze-and-Excitation (PCSE)
The standard basic block in RT-DETR uses a stack of full convolutions, leading to high GFLOPs. We propose PCSE, which leverages Partial Convolution (PConv) to update only a fraction of input channels, and then applies Squeeze-and-Excitation (SE) attention to re-calibrate all channels. The memory access of PConv is given by:
\[
\text{MAC}_{\text{PConv}} \approx 2 H W C k^2 \times \frac{c_p}{C}
\]
where \(c_p\) is the number of channels processed by convolution. Setting \(c_p/C = 1/4\) achieves a good balance between accuracy and efficiency. The PCSE block first reduces channels with a 1×1 conv, applies a 3×3 conv and then PConv, and finally adds a residual connection. SE attention then adaptively weights the channels. A hyperparameter study is shown below.
| \(c_p / C\) | mAP50 (%) | Params (M) | GFLOPs | Model Size (MB) |
|---|---|---|---|---|
| 1/2 | 45.1 | 15.3 | 45.7 | 29.7 |
| 1/4 | 44.9 | 14.1 | 42.9 | 27.5 |
| 1/6 | 44.7 | 13.9 | 42.3 | 27.1 |
| 1/8 | 44.3 | 13.8 | 42.2 | 26.9 |
Spatial-to-Channel Feature Reuse (SCFR)
Small objects lose their discriminative features as they propagate through deep layers. The SCFR module takes the shallow feature map S2 (160×160 resolution) from the backbone and transforms its spatial neighborhood information into channel dimensions to compensate for the missing detail in the small object detection layer. The process involves a non-parametric dimension conversion and a parametric feature extraction. With a scale factor of 2, adjacent pixels are rearranged into 4× channels at half spatial size. Then a 1×1 conv reduces channels, and a 3×3 dilated convolution (dilation=2) expands the receptive field. The final feature is concatenated with the original detection layer feature.
\[
\begin{aligned}
S_{21} &= \text{Concat}\big[S_{2}^{(0,0)}, S_{2}^{(0,1)}, S_{2}^{(1,0)}, S_{2}^{(1,1)}\big] \\
S_{22} &= \text{Concat}\big[\text{DCBR}_{3\times3, d=2}(\text{CBR}_{1\times1}(S_{21})),\; \text{CBR}_{1\times1}(S_{21})\big]
\end{aligned}
\]
Ablation studies confirm that using SCFR on the S2 feature improves mAP50 by 0.6% over the baseline and provides the best performance among different feature sources and fusion methods.
| Feature Source | Method | Scale/K | Fusion Style | mAP50 (%) |
|---|---|---|---|---|
| S1 (320×320) | SCFR | 4 | Direct concat | 45.9 |
| S2 (160×160) | SCFR | 2 | Direct concat | 46.2 |
| S3 (80×80) | SCFR | 1 | Direct concat | 45.8 |
| S3 (80×80) | None | \ | Direct concat | 45.6 |
| S2 (160×160) | MaxPool | 2 | Direct concat | 46.0 |
| S2 (160×160) | AvgPool | 2 | Direct concat | 45.9 |
| S2 (160×160) | SCFR | 2 | Expand add | 45.7 |
| S2 (160×160) | SCFR | 2 | Split add + concat | 45.2 |
| S2 (160×160) | SCFR | 2 | Split add + weighted | 45.4 |
Layer Adaptive Magnitude-based Pruning (LAMP)
To further reduce computational overhead, we apply LAMP pruning to the integrated model. LAMP computes a score for each weight based on its magnitude relative to the sum of larger weights in the same layer, then prunes the lowest-scoring weights. The pruning process is iterative and followed by fine-tuning. We compare global pruning with selective pruning of different detection layers. The best result is achieved by pruning only the medium and large object detection layers (speedup 1.1), resulting in a 14.4% reduction in model size without sacrificing mAP50.
| Pruning Target | Speed up | P (%) | R (%) | mAP50 (%) | Params (M) | GFLOPs | FPS (f/s) | Model Size (MB) |
|---|---|---|---|---|---|---|---|---|
| None | 1.0 | 59.9 | 44.8 | 46.0 | 14.2 | 45.5 | 94.3 | 27.7 |
| Global | 1.1 | 59.0 | 44.6 | 45.9 | 13.6 | 40.9 | 95.2 | 26.5 |
| Global | 1.2 | 57.7 | 43.2 | 44.5 | 12.7 | 35.6 | 98.0 | 24.9 |
| Global | 1.3 | 57.6 | 42.9 | 44.0 | 12.3 | 33.6 | 100.0 | 24.0 |
| Global | 1.4 | 58.5 | 40.2 | 42.6 | 12.0 | 32.5 | 103.1 | 23.4 |
| Global | 1.5 | 53.6 | 38.3 | 38.7 | 11.7 | 29.8 | 106.4 | 22.6 |
| Small + Medium | 1.1 | 59.1 | 43.9 | 45.4 | 13.3 | 41.1 | 93.5 | 26.1 |
| Small + Large | 1.1 | 58.5 | 43.6 | 45.0 | 13.2 | 40.8 | 87.7 | 25.9 |
| Medium + Large (Ours) | 1.1 | 60.6 | 44.6 | 46.1 | 12.1 | 41.3 | 97.1 | 23.7 |
Experiments and Results
Datasets and Evaluation Metrics
We evaluate our method on two drone-captured datasets: VisDrone2019 and HIT-UAV. VisDrone2019 contains 8,599 images from 14 Chinese cities, covering 10 categories (pedestrian, people, bicycle, car, van, truck, tricycle, awning-tricycle, bus, motor). HIT-UAV consists of 2,898 infrared thermal images with 5 categories (person, car, bicycle, othervehicle, dontcare). In both datasets, small objects (area < 32² pixels) dominate, as shown in the following distribution.
| Dataset | Small Objects (%) | Medium Objects (%) | Large Objects (%) |
|---|---|---|---|
| VisDrone2019 | 62.3 | 28.7 | 9.0 |
| HIT-UAV | 55.1 | 30.4 | 14.5 |
We report Precision (P), Recall (R), mAP50, and mAP50-95. Frames per second (FPS) is measured on an NVIDIA RTX 3080 Ti GPU. All models are trained for 200 epochs with batch size 4 and initial learning rate 0.0001. For LAMP pruning, we use 10 pruning iterations and 50 fine-tuning epochs.
Comparison with State-of-the-Art on VisDrone2019
We compare LA-DETR with a wide range of detectors, including R-CNN series, SSD, RTMDet, YOLO variants (v5-s/m, v6-n, v7-tiny, v8-s/m, v9-s, v10-s, v11-s, v12-s/m, v13-s, YOLO26-s/m), Mamba YOLO, MM-DETR, MobileMamba, DETR, and RT-DETR. The comprehensive results are summarized in the table below.
| Model | P (%) | R (%) | mAP50 (%) | Params (M) | GFLOPs | FPS (f/s) | Model Size (MB) |
|---|---|---|---|---|---|---|---|
| Faster R-CNN | 46.1 | 31.9 | 34.7 | 28.3 | 126.0 | 20.0 | 218.0 |
| Cascade R-CNN | 49.6 | 33.7 | 37.7 | 56.1 | 154.0 | 24.0 | 431.0 |
| SSD (MobileNetv2) | 38.2 | 31.8 | 29.9 | 3.2 | 13.8 | 71.4 | 32.5 |
| SSD (VGG-16) | 45.1 | 34.1 | 35.7 | 25.0 | 271.0 | 21.1 | 262.0 |
| RTMDet | 49.3 | 37.3 | 38.6 | 8.9 | 29.5 | 71.0 | 152.0 |
| YOLOv5-s | 45.5 | 33.9 | 33.4 | 7.0 | 15.8 | 111.1 | 13.7 |
| YOLOv5-m | 47.6 | 36.0 | 35.7 | 20.9 | 48.0 | 84.0 | 40.2 |
| YOLOv6-n | 42.9 | 34.8 | 31.7 | 4.6 | 11.3 | 153.8 | 9.9 |
| YOLOX-s | 77.0 | 25.7 | 37.8 | 8.9 | 26.8 | 66.5 | 34.3 |
| YOLOX-m | 75.8 | 32.2 | 43.6 | 25.3 | 73.8 | 58.6 | 96.8 |
| YOLOv7-tiny | 50.2 | 37.4 | 36.1 | 6.0 | 13.1 | 85.5 | 11.7 |
| YOLOv8-s | 50.8 | 39.5 | 40.3 | 11.1 | 28.7 | 128.2 | 21.4 |
| YOLOv8-m | 51.8 | 41.3 | 41.4 | 25.8 | 78.7 | 76.3 | 49.6 |
| YOLOv9-s | 49.4 | 38.1 | 38.8 | 7.2 | 26.7 | 125.0 | 14.5 |
| YOLOv10-s | 49.6 | 37.5 | 38.4 | 8.0 | 24.5 | 144.9 | 15.7 |
| YOLOv11-s | 50.4 | 38.9 | 39.7 | 9.4 | 21.3 | 131.6 | 18.3 |
| YOLOv12-s | 51.6 | 38.9 | 40.0 | 9.2 | 21.2 | 107.5 | 18.0 |
| YOLOv12-m | 52.3 | 41.8 | 42.6 | 20.1 | 67.2 | 81.4 | 38.8 |
| YOLOv13-s | 46.9 | 36.3 | 36.8 | 9.0 | 20.7 | 153.8 | 17.7 |
| YOLO26-s | 44.9 | 38.4 | 39.2 | 9.5 | 20.5 | 158.7 | 19.3 |
| YOLO26-m | 54.9 | 43.5 | 44.9 | 20.4 | 67.9 | 112.4 | 42.0 |
| Mamba YOLO-t | 46.1 | 35.9 | 36.1 | 5.7 | 12.3 | 84.0 | 11.1 |
| Mamba YOLO-m | 50.1 | 38.3 | 39.3 | 20.5 | 44.5 | 48.1 | 39.4 |
| MM-DETR-t | 50.4 | 34.3 | 34.1 | 9.1 | 17.2 | 39.2 | 17.6 |
| MM-DETR-m | 56.1 | 40.7 | 41.9 | 23.7 | 48.4 | 25.5 | 45.3 |
| MobileMamba-B | 50.8 | 43.1 | 43.8 | 27.1 | 151.0 | 38.6 | 109.0 |
| DETR | 32.6 | 24.1 | 24.0 | 28.8 | 39.4 | 54.2 | 564.0 |
| RT-DETR | 60.9 | 44.0 | 45.5 | 19.9 | 57.0 | 90.9 | 38.5 |
| LA-DETR (Ours) | 60.6 | 44.6 | 46.1 | 12.1 | 41.3 | 97.1 | 23.7 |
Our LA-DETR achieves the highest mAP50 among all compared methods while maintaining low parameter count and computational cost. Compared to RT-DETR, we reduce parameters by 39.2%, GFLOPs by 27.5%, and model size by 38.4%, while simultaneously improving mAP50 by 0.6%. The FPS of 97.1 on an RTX 3080 Ti demonstrates real-time capability suitable for drone technology.
Ablation Study
To assess the individual contributions of each module, we conduct ablation experiments by incrementally adding PCSE, CHMFA, and SCFR to the baseline RT-DETR. Results on VisDrone2019 are shown below.
| PCSE | CHMFA | SCFR | P (%) | R (%) | mAP50 (%) | mAP50-95_S | mAP50-95_M | mAP50-95_L | Params (M) | GFLOPs | Model Size (MB) |
|---|---|---|---|---|---|---|---|---|---|---|---|
| ✗ | ✗ | ✗ | 60.9 | 44.0 | 45.5 | 17.1 | 35.1 | 39.7 | 19.9 | 57.0 | 38.5 |
| ✓ | ✗ | ✗ | 59.2 | 44.3 | 44.9 | 17.0 | 34.6 | 39.1 | 14.1 | 42.9 | 27.5 |
| ✗ | ✓ | ✗ | 60.8 | 45.1 | 46.6 | 18.1 | 36.2 | 39.7 | 19.9 | 58.5 | 38.6 |
| ✗ | ✗ | ✓ | 60.5 | 44.9 | 46.2 | 17.6 | 35.5 | 39.9 | 20.0 | 58.1 | 38.7 |
| ✓ | ✓ | ✗ | 59.7 | 44.6 | 45.5 | 17.4 | 35.5 | 40.9 | 14.1 | 44.4 | 27.6 |
| ✓ | ✗ | ✓ | 59.6 | 44.5 | 45.3 | 17.6 | 34.8 | 41.6 | 14.2 | 44.0 | 27.7 |
| ✓ | ✓ | ✓ | 59.9 | 44.8 | 46.0 | 17.9 | 35.3 | 42.1 | 14.2 | 45.5 | 27.7 |
When PCSE is used alone, the model becomes significantly lighter (params drop to 14.1M) but suffers a minor mAP50 drop (0.6%). Adding CHMFA recovers the accuracy while keeping parameters low. SCFR further boosts small-object detection (mAP50-95_S increases by 0.5%). The full combination achieves a 0.5% mAP50 improvement over RT-DETR with a 28.6% reduction in parameters, demonstrating the synergistic effect of our designs.
Generalization on HIT-UAV Dataset
We further verify the generalization ability of LA-DETR on the HIT-UAV infrared dataset. Results are summarized below.
| Model | P (%) | R (%) | mAP50 (%) |
|---|---|---|---|
| Faster R-CNN | 85.7 | 55.0 | 63.5 |
| Cascade R-CNN | 80.2 | 58.1 | 68.5 |
| SSD (MobileNetv2) | 64.8 | 53.2 | 60.3 |
| SSD (VGG-16) | 77.5 | 60.7 | 68.2 |
| RTMDet | 86.9 | 73.9 | 79.3 |
| YOLOv5-s | 88.6 | 74.4 | 79.1 |
| YOLOv5-m | 92.2 | 74.3 | 79.6 |
| YOLOv6-n | 72.6 | 69.7 | 71.8 |
| YOLOX-s | 84.2 | 75.0 | 78.3 |
| YOLOv7-tiny | 83.4 | 74.0 | 78.4 |
| YOLOv8-s | 86.1 | 77.4 | 78.8 |
| YOLOv8-m | 89.6 | 77.4 | 81.8 |
| YOLOv9-s | 83.9 | 77.9 | 81.6 |
| YOLOv10-s | 84.5 | 77.4 | 80.6 |
| YOLOv11-s | 86.9 | 75.5 | 81.6 |
| YOLOv12-s | 83.1 | 78.8 | 81.0 |
| YOLOv13-s | 83.6 | 67.5 | 75.1 |
| YOLO26-s | 85.0 | 70.3 | 76.4 |
| YOLO26-m | 81.6 | 74.7 | 81.0 |
| Mamba YOLO-t | 81.0 | 73.1 | 77.6 |
| Mamba YOLO-m | 84.2 | 74.3 | 79.9 |
| MM-DETR-t | 75.2 | 69.2 | 74.7 |
| MM-DETR-m | 73.7 | 75.5 | 79.8 |
| MobileMamba-B | 63.5 | 62.0 | 67.5 |
| DETR | 69.3 | 62.3 | 65.7 |
| RT-DETR | 92.0 | 77.0 | 85.0 |
| LA-DETR (Ours) | 94.2 | 77.8 | 85.6 |
LA-DETR outperforms all competitors on the infrared modality, achieving 85.6% mAP50, which is 0.6% higher than RT-DETR and significantly better than other lightweight models. This confirms that our proposed modules generalize well across different imaging conditions (visible and infrared), a crucial advantage for drone technology applications operating in diverse environments.
Qualitative Results
Visual comparisons on sample images from VisDrone2019 and HIT-UAV datasets indicate that LA-DETR significantly reduces false detections and missed detections compared to RT-DETR. In cluttered scenes with many small objects, our model correctly identifies more true positives while suppressing spurious detections caused by reflections or background clutter.
Conclusion
In this work, we presented LA-DETR, a lightweight and accurate Transformer-based detector specifically designed for small object detection in drone technology. By introducing the CHMFA module to preserve multi-grained context, the PCSE block to reduce computational cost, the SCFR module to recover spatial details, and the LAMP pruning strategy to eliminate redundancy, we achieved a substantial reduction in parameters (39.2%) and GFLOPs (27.5%) while improving detection accuracy (0.6% mAP50 on both VisDrone2019 and HIT-UAV). Our model runs at 97.1 FPS on a consumer GPU, meeting the real-time requirements of drone deployment. Future work will focus on knowledge distillation to further boost accuracy, and we plan to deploy the algorithm on embedded platforms such as Jetson Nano to validate its practical performance in real-world drone missions. The lightweight and robust nature of LA-DETR makes it a strong candidate for next-generation drone-based detection systems.