Master Thesis Intern

Electronic Arts creates next-level entertainment experiences that inspire players and fans around the world. Here, everyone is part of the story. Part of a community that connects across the globe. A place where creativity thrives, new perspectives are invited, and ideas matter. A team where everyone makes play happen. Master thesis projects at EA Sports TRACAB - Spring 2027 The following master’s thesis projects are offered by the Computer Vision and Machine Learning team at EA SPORTS TRACAB in Stockholm. We develop computer vision and machine learning technology for capturing, reconstructing, and understanding sports. Students will work closely with our engineers on challenging, open-ended research problems using real-world sports data. The projects are expected to run from February to June 2027. The precise scope and methodology will be agreed with the selected student and may be adapted to their background, interests, and research ideas. For more information, please visit TRACAB - Official Website . To apply for one or more of the projects, please email gloy@ea.com and include: A short letter indicating which project or projects you are interested in and why. A copy of your CV. Intelligent Quality Assessment of Sports Tracking Data Modern sports tracking systems generate large volumes of complex player and ball trajectory data. Assessing the quality of this data manually is time-consuming, while simple rule-based checks may fail to identify more subtle problems. This project will investigate whether recent machine learning and foundation-model approaches can automatically identify, localize, and explain potentially unreliable segments in sports tracking data. The student will develop a rigorous evaluation framework and compare modern approaches with strong statistical and machine learning baselines. The goal is to explore systems that can provide more than a single quality score: they should also offer useful information about when data should be investigated and what may have occurred. This project is particularly suitable for students interested in foundation models, anomaly detection, time-series learning, and explainable AI. Learning Sports Events from Player and Ball Tracking Data Modern sports video and tracking systems contain rich information about the events and interactions occurring throughout a match. Recent progress in vision-language models has also demonstrated increasingly strong capabilities for understanding complex sporting scenes directly from images and video. This project will investigate modern machine learning approaches for automatically recognizing football events from visual data, player and ball tracking data, or combinations of the two. The student will work with real football footage, accurate tracking data, and corresponding event annotations. The project may explore vision-language models for direct event understanding from images or short video sequences, alongside models that reason over multivariate spatiotemporal player and ball trajectories. A key research question will be how these different representations compare in terms of accuracy, generalization, and the amount of labelled data required. The project offers scope to investigate recent advances in vision-language models, video understanding, temporal modelling, and multimodal learning. Reference: Jiayuan Rao et al., “Towards Universal Soccer Video Understanding,” CVPR 2025. The work introduces SoccerReplay-1988 and MatchVision, a soccer-specific vision-language foundation model supporting tasks including action classification and commentary generation. Open Access CVF Robust 3D Human Reconstruction for Close-Contact Sports Recent methods for human pose estimation and 3D human mesh recovery have achieved impressive results from images and video. However, close-contact sports remain especially challenging because athletes move rapidly, adopt unusual poses, physically interact, and frequently occlude one another. This project will investigate state-of-the-art 3D human reconstruction methods using multi-view footage from close-contact sports. The initial phase will establish the capabilities and limitations of recent models under these difficult conditions. The student will then investigate methods for improving reconstruction accuracy and temporal consistency, particularly during severe occlusion and physical interaction. The project is well suited to students interested in cutting-edge computer vision, human motion understanding, multi-view learning, and 3D reconstruction. Learning-Assisted Multi-Camera Calibration Accurate camera calibration is essential for reconstructing sporting activity in three dimensions from multiple video cameras. Classical geometric computer vision provides powerful methods for calibration, while recent advances in machine learning create opportunities to extract additional information from images and use it to improve calibration accuracy, robustness, or automation. This project will investigate how learning-based techniques can complement geometric calibration and numerical optimization. Using real multi-camera sports data, the student will develop and evaluate methods that extract useful scene information and incorporate it into a principled calibration framework. The project is particularly suitable for students who enjoy combining modern machine learning with multi-view geometry, mathematical modelling, and optimization. Physics-Aware 3D Ball Tracking Accurately reconstructing the trajectory of a fast-moving football in three dimensions is a challenging computer vision problem. The ball may be small in the image, move rapidly, suffer from motion blur, become temporarily occluded, or be visible from only a subset of cameras. At the same time, its motion is constrained by physical behaviour that can provide valuable information when visual observations are noisy or missing. This project will investigate machine learning methods that combine visual evidence, tempora