BROWSE
Detail
Conference Paper
An Improved YOLOF for Scale Imbalance with Dilated Attention
Cited 0 time in 
Share 
Share
- Authors
- Tsatsral Amarbayasgalan, Mooseop Kim, Chi Yoon Jeong
- Issue Date
- 2024-12
- Citation
- International Conference on Pattern Recognition (ICPR) 2024 (LNCS 15317), pp.156-172
- Publisher
- Springer
- Language
- English
- Type
- Conference Paper
- Abstract
- Scale imbalance, where objects of different sizes are not equally represented in a dataset, is a common problem in real-world object detection scenarios that leads to significant performance degradation of object detection methods. Although several solutions have been proposed based on multilevel feature maps, these methods may not be suitable in real-time applications owing to their low speed and memory consumption. Recently, you only look one-level feature (YOLOF) was proposed based on a single-in-single-out (SiSo) architecture; the SiSo architecture is well suited for real-time applications with performance comparable to that of methods based on multilevel feature maps. However, they show limited performance when applied to real-world object detection scenarios with scale imbalance problems. Therefore, we propose a lightweight object detection method that can handle the scale imbalance problem while retaining the advantages of the SiSo framework. To mitigate the scale imbalance, we use dilated attention to extend the SiSo architecture and learn the scale range of objects. Extensive experiments on public datasets show the effectiveness of a dilated attention-based proposed method in scale-imbalanced scenarios. Our method achieves results comparable to those of the original YOLOF on the MS COCO and PASCAL VOC datasets. In particular, for imbalanced datasets, the proposed method outperforms the original YOLOF by 4.78% on the first-person-walking-livingroom dataset and by 1.38% on the imbalanced PASCAL VOC dataset in terms of average precision (AP)50.
- KSP Keywords
- Average Precision, Common problem, Detection Method, Different sizes, Feature map, First-person, Imbalance Problem, Imbalanced datasets, Low speed, PASCAL VOC dataset, Public Datasets