UFO-DETR: Frequency-Guided End-to-End Detector for UAV Tiny Objects
Authors: Yuankai Chen, Kai Lin, Qihong Wu, Xinxuan Yang, Jiashuo Lai, Ruoen Chen, Haonan Shi, Minfan He, +1 more
Organizations: 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 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.
Figures & tables
Fig. 1: Overview of the UFO-DETR. It is divided into the backbone, encoder, and detection head. The backbone is rebuilt using LSKNet. The decoder is implemented through CCFD based on DyFreqC3 design and AIFI with integrated deformable attention. The detection head consists of the decoder and auxiliary prediction heads, completing the final object detection task.
Fig. 2: Structure of the LSKNet Block and LK Selection.
Fig. 3: An illustration of Deformable Attention mechanism.
Fig. 5: Visual comparison results of RTDETR-L, YOLOv8M, YOLOv10, YOLOv11M. (a) shows Repeated Detections of the same object, (b) corresponds to model false positives, and (c) denotes objects that the model failed to detect.
Model
P/%
R/%
mAP50/%
GFLOPs
Model Size/MB
baseline
59.0
42.4
43.5
103.5
66.2
baseline + A
58.1
41.7
43.4
37.6
26.0
baseline + A + B
59.0
43.0
44.2
37.8
26.0
baseline + A + B + C
59.2
44.5
46.1
41.8
28.3
TABLE I: Results of the ablation study.
Model
P/%
R/%
mAP50/%
GFLOPs
Model Size/MB
YOLOv8-M
53.5
37.4
40.7
78.9
52
YOLOv8-L
51.8
39.7
40.6
164.9
87.7
YOLOv10-M
49.9
36.7
37.9
63.5
33.5
YOLOv10-L
51.6
38.4
39.8
126.4
52.2
YOLOv11-M
51.8
38.1
39.6
67.7
40.5
YOLOv11-L
53.0
38.3
39.8
86.6
51.2
TABLE II: Comparison Results with Other Object Detectors.
Fig. 6: The heatmap of RTDETR-L and UFO-DETR. The brighter areas in the heatmap indicate stronger attention by the model.
College of Computer Science and Technology, Zhejiang Normal University, Jinhua, China · Hangzhou School of Automation, Zhejiang Normal University, Hangzhou, China · Beijing Geekplus Technology Co., Ltd, Beijing, China
College of Computer and Information Science, Chongqing Normal University, Chongqing 401333, China · CETC Yizhihang (Chongqing) Technology Co., Ltd, Chongqing 400031, China · School of Computer Engineering, Weifang University, Weifang 261061, China +1