Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

MGDFIS: Multi-scale Global-detail Feature Integration Strategy for Small Object Detection

Created by
  • Haebom

Author

Yuxiang Wang, Xuecheng Bai, Boyu Hu, Chuanzhi Xu, Haodong Chen, Vera Chung, Tingxue Li, Xiaoming Chen

Outline

This paper proposes a novel multi-scale global-detail feature integration strategy (MGDFIS) for small object detection in unmanned aerial vehicle (UAV) images. To address the computational overhead and blurry details of existing multi-scale fusion methods, MGDFIS employs an integrated fusion framework that tightly combines global context and local detail. Composed of three modules (FusionLock-TSS Attention Module, Global-detail Integration Module, and Dynamic Pixel Attention Module), MGDFIS enhances small object detection performance by emphasizing spectral and spatial cues, efficiently integrating multi-scale context, and rebalancing imbalanced foreground and background distributions. Experimental results on the VisDrone benchmark demonstrate that MGDFIS outperforms state-of-the-art methods across a variety of backbone architectures and detection frameworks.

Takeaways, Limitations

•
Takeaways:
◦
We present an efficient fusion framework that significantly improves small object detection performance in UAV images.
◦
A balanced design that maximizes accuracy while minimizing computational cost.
◦
Demonstrated excellent performance across various backbone architectures and detection frameworks.
◦
Providing practical solutions suitable for resource-constrained UAV platforms.
•
Limitations:
◦
Further validation of the proposed method's generalization performance is needed (although experimental results on other datasets are lacking).
◦
There is a possibility that it may exhibit biased performance on certain types of small objects (lack of performance analysis across various object types).
◦
Lack of experimental results in real-world UAV environments (simulated data or limited availability of real-world data).
👍
Made with Slashpage