Related Experiment Video
Updated: Jun 17, 2025
03:31
Author Spotlight: Enhancement of Salient Object Detection for Smart Grid Applications
Published on: December 15, 2023
499
Global remote feature modulation end-to-end detection.
XiaoAn Bao1, WenJing Yi1, XiaoMei Tu2
1School of Computer Science and Technology, Zhejiang Sci-Tech University, Hangzhou, 310018, Zhejiang Province, China.
Scientific Reports
|August 7, 2024
Summary
This study introduces a Global Remote Feature Modulation End-to-End (GRFME2E) algorithm to improve object detection accuracy, especially in dense and occluded scenes. The GRFME2E algorithm achieves high accuracy on pig and crowd datasets.
Area of Science:
- Computer Vision
- Deep Learning
- Object Detection
Background:
- Fully convolutional networks excel in object detection but struggle with dense and occluded targets.
- Existing algorithms lack accuracy in challenging dense and occluded scenarios.
Purpose of the Study:
- To propose a novel Global Remote Feature Modulation End-to-End (GRFME2E) detection algorithm.
- To address the limitations of current object detectors in dense and occluded scenes.
Main Methods:
- Introduced Concentric Attention Feature Pyramid Network (CAFPN) for enhanced feature extraction, capturing global dependencies and position-sensitive information.
- Developed a Two-Stage Detection Head (TS Head) with First-One-to-Few (F-O2F) and Second-One-to-Few (S-O2F) modules to handle unobstructed and occluded objects respectively.
- Integrated CAFPN and TS Head for end-to-end object detection.
Main Results:
- Achieved 98.4% accuracy on a pig detection dataset.
- Attained 91.8% accuracy on the CrowdHuman dataset, outperforming existing methods.
- Demonstrated effective detection of objects across varying degrees of occlusion.
Conclusions:
- The GRFME2E algorithm significantly improves object detection performance in challenging dense and occluded environments.
- CAFPN and TS Head effectively enhance feature representation and detection capabilities.
- GRFME2E offers a robust solution for real-world object detection applications.
