Related Experiment Video
Updated: Jul 5, 2025
04:23
A Swin Transformer-Based Model for Thyroid Nodule Detection in Ultrasound Images
Published on: April 21, 2023
1.9K
An enhanced Swin Transformer for soccer player reidentification
Sara Akan1, Songül Varlı1, Mohammad Alfrad Nobel Bhuiyan2
1Department of Computer Engineering, Yildiz Technical University, Istanbul, Turkey.
Scientific Reports
|January 11, 2024
Summary
This study introduces a new method for re-identifying players in sports videos, improving accuracy even with similar uniforms and low-resolution images. The enhanced Swin Transformer model achieves state-of-the-art results on benchmark datasets.
Area of Science:
- Computer Vision
- Machine Learning
- Sports Analytics
Background:
- Object re-identification (ReID) is crucial for applications like sports analytics.
- Existing methods struggle with similar team attire, limited samples, and low image resolution in sports videos.
- Convolutional Neural Network (CNN) and Transformer models have limitations in capturing long-range dependencies and small object features.
Purpose of the Study:
- To develop an improved ReID method for identifying players in broadcast sports videos.
- To address challenges including similar clothing, few samples per player, and low image resolution.
- To enhance feature representation by combining CNNs and Transformers for robust object ReID.
Main Methods:
- Developed a Regional Feature Extraction Swin Transformer (RFES) backbone to improve local and small-scale object feature extraction.
- Utilized three loss functions to manage imbalanced data and emphasize difficult cases.
- Implemented re-ranking with k-reciprocal encoding in the retrieval phase.
Main Results:
- Achieved rank-1 accuracy of 96.2% with mAP of 89.1 on the Market-1501 dataset.
- Reached rank-1 accuracy of 84.1% with mAP of 86.7 on the SoccerNet-v3 dataset.
- Outperformed existing state-of-the-art approaches on both datasets.
Conclusions:
- The proposed RFES method effectively addresses the challenges of player ReID in sports videos.
- Combining CNNs and Transformers within the Swin Transformer architecture enhances feature representation.
- The method demonstrates superior performance, setting a new benchmark for sports player re-identification.
