S-amba: A Multi-View Foul Recognition in Soccer Through a Mamba-Based Approach

Document Type

Conference Proceeding

Publication Date

1-1-2025

Abstract

In this work, we propose a novel Mamba-based multi-task framework for multi-view foul recognition. Our approach leverages the Mamba architecture's efficient long-range dependency modeling to process synchronized multi-view video inputs, enabling robust foul detection and classification in soccer matches. By integrating spatial-temporal feature extraction with a multi-task learning strategy, our model simultaneously predicts foul occurrences, identifies foul types, and localizes key events across multiple camera angles. We employ a hybrid loss function to balance classification and localization objectives, enhancing performance on diverse foul scenarios. Extensive experiments on the SoccerNet-MVFoul dataset demonstrate our method's superior accuracy and efficiency compared to traditional CNN and Transformer-based models. Our framework achieves competitive results, offering a scalable and real-time solution for automated foul recognition, advancing the application of computer vision in sports analytics. The codebase is publicly available at https://github.com/areyesan/Mamba-Based MVFR for reproducibility.

Publication Title

International Conference on Sport Sciences Research and Technology Support Icsports Proceedings

ISBN

[9789897587719]

Link to Full Text

Share

COinS