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.

Exponentially Weighted Instance-Aware Repeat Factor Sampling for Long-Tailed Object Detection Model Training in Unmanned Aerial Vehicles Surveillance Scenarios

Created by
  • Haebom

Author

Taufiq Ahmed, Abhishek Kumar, Constantino Alvarez Casado, Anlan Zhang, Tuomo H anninen, Lauri Loven, Miguel Bordallo L opez, Sasu Tarkoma

Outline

This paper proposes Exponentially Weighted Instance-Aware Repeat Factor Sampling (E-IRFS), an improvement on existing sampling-based rebalancing strategies (RFS, IRFS) to address the class imbalance problem commonly encountered in object detection models. E-IRFS uses exponential weighting, which applies an exponential function to the geometric mean of image and instance frequencies, to better reflect the differences between rare and frequent classes. Experiments conducted on the Fireman-UAV-RGBT dataset and four other publicly available datasets, using the YOLOv11 model, show that E-IRFS improves rare class detection performance by 22% compared to existing methods, particularly on lightweight models with limited capacity. This contributes to improving rare object detection performance in resource-constrained environments and demonstrates its suitability for real-time applications such as UAV-based emergency monitoring. The source code is available on GitHub.

Takeaways, Limitations

•
Takeaways:
◦
We demonstrate that E-IRFS using exponential weighting is more effective in solving class imbalance problems with long-tailed distributions than conventional linear weight-based sampling techniques.
◦
It has shown great effectiveness in solving class imbalance problems, especially in lightweight models, and has contributed to improving real-time object detection performance in resource-constrained environments.
◦
Presenting applicability in real-world applications such as UAV-based emergency monitoring.
◦
Ensure reproducibility and extensibility through open source code.
•
Limitations:
◦
Since the experiments were conducted using limited datasets (Fireman-UAV-RGBT dataset and four public datasets), verification of generalization performance on other datasets is necessary.
◦
Since the experimental results are limited to the YOLOv11 model, performance evaluation for other object detection models is required.
◦
Lack of detailed discussion on parameter optimization of the exponential function of E-IRFS.
👍
Made with Slashpage