MITE-Net: SWaP-Optimized 4K Video Tiny Target Perception for Embodied Edge SAR

Mingshuo Xu; Mu Hua; Jigen Peng; Qi Wang; Shigang Yue

cs.CVAug 16, 2026

MITE-Net: SWaP-Optimized 4K Video Tiny Target Perception for Embodied Edge SAR

Authors: Mingshuo Xu, Mu Hua, Jigen Peng, Qi Wang, Shigang Yue

Organizations: School of Mathematics and Computing Science, University of Leicester, Leicester LE1 7RH, UK · Machine Life and Intelligence Research Center, Guangzhou University, Guangzhou 510006, China

Abstract

Real-time tiny target perception in high-resolution imagery is critical for embodied Search-and-Rescue (SAR) missions. However, strict Size, Weight, and Power (SWaP) constraints on edge devices like UAVs create a bottleneck: traditional image downsampling causes severe feature loss, while slice-based processing incurs prohibitive latency. To address this gap, this paper introduces a comprehensive framework encompassing a novel architecture, specialized datasets, and hardware-level benchmarks. First, we propose MITE-Net, a SWaP-optimized cascaded architecture, which couples a bio-inspired, learning-free Tiny Target Motion-Based Region Proposal Network (TTM-RPN) with a sub-0.14M-parameter R-CNN-like head. Second, to standardize 4K tiny target evaluation, we construct the SAR-Tiny Datasets by relabeling two challenging UAV datasets: SeaDroneSee-Tiny (dynamic maritime scenes, tiny targets predominantly of 64-256 pixels ) and UAVID-Tiny (cluttered urban scenes, extremely tiny targets, less than 64 pixels). Third, we benchmark against state-of-the-art YOLO models on an edge device, NVIDIA Jetson AGX Xavier, where MITE-Net directly processes 4K maritime imagery, achieving a 100% search success rate at 30.33 FPS. Consuming merely 3.19 W (9.51 FPS/W), MITE-Net vastly outperforms YOLO baselines in target recall and energy efficiency. Conversely, UAVID-Tiny evaluations expose a compound structural limitation: the learning-free bionic front-end struggles against urban backgrounds, while the ultra-lightweight head lacks representational capacity for complex features. Ultimately, this work delivers an efficient onboard perception paradigm and a rigorous baseline guiding future end-to-end SAR architectures.

Figures & tables

Fig. 1: Comparison of search success rate versus processing speed (FPS) among different models. The green shaded area denotes the real-time processing region ( $\geq$ 30 FPS), and the annotated numbers indicate the efficiency metric for each tested configuration. While YOLOv8n-P2 achieves a 100% success rate, it is limited to suboptimal speeds ( $<$ 5 FPS). Conversely, YOLOv11n approaches real-time speed but suffers a significant drop in success rate. Our proposed MITE-Net achieves an optimal trade-off, successfully maintaining a 100% search success rate within the real-time processing region while demonstrating significantly higher efficiency (12.9 $\times$ ).

Fig. 2: Qualitative demonstration of 4K tiny target perception for the SAE mission using the proposed MITE-Net and baseline models. The top row displays the original 4K resolution images, while the bottom panels show the corresponding zoomed-in regions. As observed, due to the extremely tiny scale of the targets, they appear merely as blurry spots even after significant magnification. The first column presents results on the SeaDroneSee-Tiny dataset, where our proposed model demonstrates excellent detection performance. The second and third columns exhibit results on the UAVID-Tiny dataset. Specifically, in the second column, our model successfully localizes a target but suffers from a classification failure. In the third column, our model experiences a bounding box regression failure when detecting multiple tiny objects. Notably, when these targets consist of merely a dozen pixels, the recall rate of mainstream YOLO baselines also drops to near zero.

Fig. 3: Overall architecture of the proposed MITE-Net. The cascaded framework consists of two stages. Left: A learning-free Bionic-Driven TTM-RPN utilizes vSTMD [ 4 ] to extract motion-salient RoIs, which are cropped and concatenated with spatial coordinates into multi-channel inputs. Right: An ultra-lightweight R-CNN-like Head (¡0.14M parameters) processes these inputs using dilated and group convolutions, followed by a dual-branch channel attention structure for precise classification and bounding box regression.

Dataset	Split	Sequences	Frames	BBoxes	Target Density	Key Characteristics
SeaDroneSee-Tiny	Train	seq 2, 3, 4, 5, 6, 7, 8	3858	7178	1-5 / frame	Dynamic maritime backgrounds.
	Val	seq 9	1001	3003	3 / frame	Baseline maritime background evaluation.
	Test	seq 1	1001	11243	12 / frame	Density mimicking real-world SAR crises.
UAVID-Tiny	Train	seq 2, 5, 7, 8, 16, 17, 33	3208	63398	10-50/ frame	Massive urban clutter.
	Val	seq 24	901	15063	$\sim$ 16 / frame	Occlusion and varied lighting conditions.
	Test	seq 23	701	9839	$\sim$ 14 / frame	Early-stage target discovery in urban scenes.

TABLE I: Detailed Sequence Allocation and Attributes of the 4K SAR-Tiny Datasets for Embodied Edge SAR

Fig. 4: Statistical overview and visual example of the curated SeaDroneSee-Tiny dataset. (A) Object area frequency distribution across the entire dataset, demonstrating a heavily right-skewed trend, where the majority of targets are categorized as Extremely Tiny (ET, Area $\leq$ 64 px 2 ) and Tiny (64 $<$ Area $\leq$ 256 px 2 ). (B) Joint distribution of target width and height across the training, validation, and test splits, partitioned by the three defined scale regimes: ET, Tiny, and Small (256 $<$ Area $\leq$ 1024 px 2 ). The highly concentrated contour of the test set in the ET and Tiny regions highlights the deliberate ”stress test” design. (C) An example frame from the dataset illustrating the visual challenges of detecting multiple scattered targets. The bounding boxes denote ground truth (GT) labels inherited from the original dataset: GT 1 represents a swimmer, and GT 2 represents a swimmer with a life jacket.

Fig. 5: Statistical overview and visual example of the curated UAVID-Tiny dataset. (A) Object area frequency distribution, highlighting a massive concentration of targets strictly within the Extremely Tiny (ET, Area $\leq$ 64 px 2 ) region. (B) Joint distribution of target width and height across data splits, where the deep penetration into the ET region underscores the dataset’s complexity. (C) Example test frame of minute targets in a cluttered urban background. The bounding boxes denote ground truth (GT) labels: GT 1 represents pedestrians, while the dataset also includes moving (GT 2) and static cars (GT 3).

			YOLOv11n [ 10 ]	YOLOv8n-P2 [ 9 ]		MITE-Net (Proposed)
			Resize $1920\times 1088$	Resize $1920\times 1088$	SAHI [ 11 ] $1344\times 768\times 9$	Raw 4K + 2ds	Raw 4K + 8ds
SAR Criteria	SDS-T	Search Success Rate (%, $\uparrow$ )	83.33	83.33	100	100	83.33
		False Alarm Rate (%, $\downarrow$ )	6.83	16.77	3.12	67.72	82.29
		Max Search Time (Frame, $\downarrow$ )	126	74	385	299	119
		Avg. Search Time (Frame, $\downarrow$ )	17.90	11.20	62.58	69.00	36.80
	U-T	Search Success Rate (%, $\uparrow$ )	11.11	22.22	36.11	8.33	0.00

TABLE II: Comprehensive Comparison of Search and Rescue (SAR) and Size, Weight, and Power (SWaP) Metrics on the SeaDroneSee-Tiny (SDS-T) and UAVID-Tiny (U-T) Datasets. Inference was conducted on an NVIDIA Jetson AGX Xavier using Float16 precision. Highlighted rows indicate primary evaluation criteria. Bold and underlined values represent the optimal and sub-optimal performances, respectively. Grayed text denotes severe algorithmic degradation or failure.

Configurations		SAR Criteria		SWaP Criteria
ds	$N$	SSR ( $\uparrow$ )	FAR ( $\downarrow$ )	FPS ( $\uparrow$ )	Eff. ( $\uparrow$ )	Total FLOPs (G, $\downarrow$ )	TTM-RPN FLOPs (G, $\downarrow$ )		R-CNN-like Head FLOPs (G, $\downarrow$ )
ds	$N$	SSR ( $\uparrow$ )	FAR ( $\downarrow$ )	FPS ( $\uparrow$ )	Eff. ( $\uparrow$ )	Total FLOPs (G, $\downarrow$ )	vSTMD	Others	R-CNN-like Head FLOPs (G, $\downarrow$ )
1	25	100.00	67.83	11.81	7.43	2.1851	1.9972	0.0317	0.1561
2	10	100.00	35.41	30.94	10.18	0.5268	0.3689	0.0258	0.1321
	25	100.00	67.72	30.33	10.32	0.5605	0.3689	0.0258	0.1657
	50	100.00	82.76	30.16	10.12	0.5607	0.3689	0.0258	0.1659

TABLE III: Model scalability of the proposed MITE-Net concerning downsampling factor (

ds

) and proposal bounding-box number (

N

) on SeaDroneSee-Tiny. SSR: Search Success Rate (%); FAR: False Alarm Rate (%); FPS: Frames Per Second; Eff.: Efficiency (FPS/Watt); FLOPs: Floating Point Operations (G). Up/down arrows indicate the preferred direction of the metric.

Module	SeaDroneSee-Tiny		UAVID-Tiny
Module	R 0.1	R 0.5	R 0.1	R 0.5
TTM-RPN (No-Category)	0.9584	0.5812	0.1047	0.0580
R-CNN-Head (Category)	0.7377	0.7072	0.0230	0.0103

TABLE IV: Module-wise performance analysis of the proposed two-stage MITE-Net.

R_{a}

denotes the recall at an NWD threshold of

a

. The significant degradation on UAVID-Tiny highlights the limitations of the ultra-lightweight head and non-learning RPN in complex urban scenes.

Motion Information	R 0.1	R 0.5
$\checkmark$	0.7377	0.7072
$\times$	0.7306	0.7012

TABLE V: Ablation study of motion feature fusion of the R-CNN-Head on SeaDroneSee-Tiny.

R_{a}

represents the recall at an NWD threshold of

a

Panel of Table 11 — TABLE VI: Hyperparameter Analysis on the SeaDroneSee-Tiny Validation Set. The table reports the $F_{2}$ score across various combinations of batch size (BS) and learning rate (LR).

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Fig. 6: Size selectivity curves of the model for target widths of (A) 4 pixels, (B) 12 pixels, and (C) 16 pixels. By configuring the suppression kernel size to correspond with the input target dimensions, the model consistently achieves its peak relative response when the target height matches the specified target width. The optimal height range yielding this peak response is highlighted by the pink shaded area (Prefer Region). The qualitative trend remains robust across various pooling parameters. Absolute values on the y-axis are omitted to emphasize the relative distribution profile.

Dataset	Model	Resolution	mAP			mAR			m $F_{2}$			FPS
Dataset	Model	Resolution	$et$	$tiny$	$all$	$et$	$tiny$	$all$	$et$	$tiny$	$all$	FPS
SeaDroneSee -Tiny	YOLOv8n-P2 [ 9 ]	SAHI [ 11 ] ( $1344\times 768\times 9$ )	0.46	0.59	0.54	0.60	0.63	0.64	0.60	0.65	0.65	4.79
	YOLOv11n [ 10 ]	Resize ( $1920\times 1088$ )	0.24	0.46	0.18	0.24	0.48	0.46	0.27	0.48	0.46	29.15
	YOLOv8n-P2 [ 9 ]	Resize ( $1920\times 1088$ )	0.26	0.42	0.28	0.26	0.49	0.47	0.28	0.47	0.37	31.22
	MITE-Net (Ours)	Raw (4K + 2ds)	0.27	0.38	0.08	0.47	0.48	0.48	0.30	0.40	0.38	30.33
UAVID-Tiny	YOLOv8n-P2 [ 9 ]	SAHI [ 11 ] ( $1344\times 768\times 9$ )	0.04	0.34	0.16	0.04	0.47	0.31	0.03	0.31	0.21	4.15

Appendix

TABLE VII: Performance comparison on SeaDroneSee-Tiny and UAVID-Tiny datasets. Rows with gray-shaded FPS denote theoretical baselines that process raw or sliced 4K imagery to establish accuracy upper bounds, though they fail to meet the real-time requirement (

\geq 30

FPS) for embodied edge deployment. In contrast, the unshaded rows represent configurations optimized for real-time edge processing.

Explore similar work

Mar 26, 2026

SDD-YOLO: A Small-Target Detection Framework for Ground-to-Air Anti-UAV Surveillance with Edge-Efficient Deployment

Detecting small unmanned aerial vehicles (UAVs) from a ground-to-air (G2A) perspective presents significant challenges, including extremely low pixel occupancy, cluttered aerial backgrounds, and strict real-time constraints. Existing YOLO-based detectors are primarily optimized for general object detection and often lack adequate feature resolution for sub-pixel targets, while introducing complexities during deployment. In this paper, we propose SDD-YOLO, a small-target detection framework tailored for G2A anti-UAV surveillance. To capture fine-grained spatial details critical for micro-targets, SDD-YOLO introduces a P2 high-resolution detection head operating at 4 times downsampling. Furthermore, we integrate the recent architectural advancements from YOLO26, including a DFL-free, NMS-free architecture for streamlined inference, and the MuSGD hybrid training strategy with ProgLoss and STAL, which substantially mitigates gradient oscillation on sparse small-target signals. To support our evaluation, we construct DroneSOD-30K, a large-scale G2A dataset comprising approximately 30,000 annotated images covering diverse meteorological conditions. Experiments demonstrate that SDD-YOLO-n achieves a [email protected] of 86.0% on DroneSOD-30K, surpassing the YOLOv5n baseline by 7.8 percentage points. Extensive inference analysis shows our model attains 226 FPS on an NVIDIA RTX 5090 and 35 FPS on an Intel Xeon CPU, demonstrating exceptional efficiency for future edge deployment.

Pengyu Chen, Haotian Sa, Yiwei Hu +2

Chien-Shiung Wu College, Southeast University, Nanjing, China · School of Information Science and Engineering, Southeast University, Nanjing, China

Small Object Detection Aerial Object Detection Real-Time Object Detection Efficient Object Detection

Ask PDF

Feb 26, 2026

UFO-DETR: Frequency-Guided End-to-End Detector for UAV Tiny Objects

Small target detection in UAV imagery faces significant challenges such as scale variations, dense distribution, and the dominance of small targets. Existing algorithms rely on manually designed components, and general-purpose detectors are not optimized for UAV images, making it difficult to balance accuracy and complexity. To address these challenges, this paper proposes an end-to-end object detection framework, UFO-DETR, which integrates an LSKNet-based backbone network to optimize the receptive field and reduce the number of parameters. By combining the DAttention and AIFI modules, the model flexibly models multi-scale spatial relationships, improving multi-scale target detection performance. Additionally, the DynFreq-C3 module is proposed to enhance small target detection capability through cross-space frequency feature enhancement. Experimental results show that, compared to RT-DETR-L, the proposed method offers significant advantages in both detection performance and computational efficiency, providing an efficient solution for UAV edge computing.

Yuankai Chen, Kai Lin, Qihong Wu +6

Easy Lab, College of Mathematics and Informatics, South China Agricultural University, Guangzhou, China · School of Computer Science and Engineering, Southeast University, Nanjing, China · School of Mathematics, Foshan University, Foshan, China

Small Object Detection Aerial Object Detection Efficient Object Detection Detection Transformer Frequency-Domain Feature Learning

Ask PDF

Apr 25, 2026

Resource-Constrained UAV-Based Weed Detection for Site-Specific Management on Edge Devices

Weeds compete with crops for light, water, and nutrients, reducing yield and crop quality. Efficient weed detection is essential for site-specific weed management (SSWM). Although deep learning models have been deployed on UAV-based edge systems, a systematic understanding of how different model architectures perform under real-world resource constraints is still lacking. To address this gap, this study proposes a deployment-oriented framework for real-time UAV-based weed detection on resource-constrained edge platforms. The framework integrates UAV data acquisition, model development, and on-device inference, with a focus on balancing detection accuracy and computational efficiency. A diverse set of state-of-the-art object detection models is evaluated, including convolution-based YOLO models (v8-v12) and transformer-based RT-DETR models (v1-v2). Experiments on three edge devices (Jetson Orin Nano, Jetson AGX Xavier, and Jetson AGX Orin) demonstrate clear trade-offs between accuracy and inference latency across models and hardware configurations. Results show that high-capacity models achieve up to 86.9% mAP50 but suffer from high latency, limiting real-time deployment. In contrast, lightweight models achieve 66%-71% mAP50 with significantly lower latency, enabling real-time performance. Among all models, RT-DETRv2-R50-M achieves competitive accuracy (79% mAP50) with improved efficiency, while YOLOv10n provides the fastest inference speed. YOLOv11s and RT-DETRv2-R50-M offer the best balance between accuracy and speed, making them strong candidates for real-time UAV deployment.

Linyuan Wang, Haibo Yao, Te-Ming Tseng +4

Department of Agricultural and Biological Engineering, Mississippi State University, Mississippi State, MS, USA · USDA-ARS, Genetics and Sustainable Agriculture Research Unit, Mississippi State, MS, USA · Department of Plant and Soil Sciences, Mississippi State University, Mississippi State, MS, USA +2

Agricultural Remote Sensing Weed Detection Edge Inference Real-Time Object Detection UAV Remote Sensing

Ask PDF