Improving Small Object Detection with Area-Scaled and Dynamic Focal Loss
Authors:
- 6 January 2025
- conference paper
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 00930-00936
Abstract
Detecting small objects remains a persistent challenge for state-of-the-art object detection models, which excel at identifying larger objects but struggle with smaller ones due to limited pixel representation and susceptibility to noise and occlusion. This paper introduces two innovative loss-scaling strategies to address these challenges: Inverse Quadratic Scaling, which adjusts loss inversely proportional to the square root of object area, and Dynamic Focal Loss, which incorporates object size into the focal loss framework. These methods aim to improve training dynamics for small objects while preserving performance for larger objects. We evaluate our strategies on the CPPES dataset using a ResNet-50-based DETR model, characterized by significant object size variations. Experimental results demonstrate substantial improvements in small object detection, achieving F1 score increases of up to 274% in some categories compared to the baseline. Notably, Inverse Quadratic Scaling excels in enhancing underrepresented objects like goggles, while Dynamic Focal Loss offers balanced performance across sizes. These findings underscore the potential of tailored loss functions to mitigate small object detection challenges and provide a foundation for real-world applications such as medical imaging, autonomous vehicles, and surveillance systems.Keywords
This publication has 17 references indexed in Scilit:
- MSF-YOLO: A multi-scale features fusion-based method for small object detectionMultimedia Tools and Applications, 2024
- An improved algorithm for small object detection based on YOLO v4 and multi-scale contextual informationComputers and Electrical Engineering, 2022
- EfficientDet: Scalable and Efficient Object Detection2020
- Deep Learning for Generic Object Detection: A SurveyInternational Journal of Computer Vision, 2019
- Mask R-CNN2017
- Feature Pyramid Networks for Object Detection2017
- Speed/Accuracy Trade-Offs for Modern Convolutional Object Detectors2017
- SSD: Single Shot MultiBox Detector2016
- You Only Look Once: Unified, Real-Time Object Detection2016
- Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation2014