Temporal Action Detection in Sports Videos


The goal of this PhD project is to investigate the task of temporally detecting actions in videos. Specifically, it involves localizing, with a certain margin of error, the temporal occurrence of actions within untrimmed videos, along with identifying the nature of these actions. Moreover, the focus is specifically on sports-related videos, characterized by their fast-paced nature.

Methods solving this task face multiple problems, such as limited amounts of data in most cases, unbalanced distributions of the different actions, with some of them occurring just a few times, and a lack of precision in their predictions. To tackle these challenges, in this project, we will explore the following areas: (1) Self-supervised learning, which can be beneficial in small datasets and addressing imbalanced data distributions, (2) few-shot learning, which focuses on learning from a minimal number of examples, and (3) token discriminability, as methodologies capable of generating more distinct feature vectors across concurrent temporal positions can improve the precision of the predictions. 

Consequently, the overall goal of this PhD is to develop methodologies that effectively address these challenges, and that can achieve SOTA results in the task of temporally detecting actions within sports videos.

Scientific Work

ASTRA: An Action Spotting TRAnsformer for Soccer Videos
Xarles, A., Escalera, S., Moeslund, T. B., & Clapés, A. (2023, October). In Proceedings of the 6th International Workshop on Multimedia Content Analysis in Sports (pp. 93-102).


This PhD program is supported by the PREDOCS-UB scholarship, jointly funded by the Universitat de Barcelona and Banco Santander.


PhD-student: Artur Xarles Esparraguera
Email: arturxe@gmail.com

Supervisor: Sergio Escalera
Email: sergio.escalera.guerrero@gmail.com

Supervisor: Albert Clapés
Email: aclapes@ub.edu

Supervisor: Thomas B. Moeslund
Email: tbm@create.aau.dk