EAST-SPL: Event-Aware Statistical Tiling for Decomposable Soccer Player Localization with an Auxiliary Rejection Network
Chaman Motlagh, Abolfazl LU
and Nilsson, Mikael
LU
(2026)
Pattern Recognition
In Springer Lecture Notes in Computer Science
16825.
p.413-428
- Abstract
- Localization of players is a fundamental step in sports analytics and automated broadcasting, yet high-performance models for 4K resolution inputs create a significant computational bottleneck for real-time processing. Recent decomposable methods, such as DTSPL-BEV, address this by processing image “tiles” but rely on static, single-frame optimization and unstable heuristic merging algorithms. In this work, we propose EAST-SPL, an Event-Aware Statistical Tiling framework that shifts the optimization paradigm from minimizing single-frame FLOPs to minimizing Expected Total FLOPs (ETF). By leveraging the temporal distribution of player locations, our objective function prioritizes granular tiling in high-probability zones while allowing... (More)
- Localization of players is a fundamental step in sports analytics and automated broadcasting, yet high-performance models for 4K resolution inputs create a significant computational bottleneck for real-time processing. Recent decomposable methods, such as DTSPL-BEV, address this by processing image “tiles” but rely on static, single-frame optimization and unstable heuristic merging algorithms. In this work, we propose EAST-SPL, an Event-Aware Statistical Tiling framework that shifts the optimization paradigm from minimizing single-frame FLOPs to minimizing Expected Total FLOPs (ETF). By leveraging the temporal distribution of player locations, our objective function prioritizes granular tiling in high-probability zones while allowing larger tiles in background regions. To realize these statistical gains, we introduce a lightweight auxiliary Rejection Network sharing convolutional features with the primary backbone to efficiently bypass empty tiles. Furthermore, we employ a Genetic Algorithm to navigate the non-differentiable configuration space, eliminating the need for heuristic merging. Experiments on the SynLoc dataset demonstrate that EAST-SPL reduces the expected computational cost by over 84% compared to conventional single-frame baselines, while also achieving substantially faster optimization than exhaustive grid search. (Code is available at https://github.com/AbolfazlChM95/EAST-SPL.) (Less)
Please use this url to cite or link to this publication:
https://lup.lub.lu.se/record/bf18201c-235e-4388-b8d2-4f472506958f
- author
- Chaman Motlagh, Abolfazl
LU
and Nilsson, Mikael
LU
- organization
- publishing date
- 2026-08-03
- type
- Chapter in Book/Report/Conference proceeding
- publication status
- published
- subject
- keywords
- Object detection, Early rejection network, Decomposable Methods, Efficient ML, Genetic Algorithm
- host publication
- Pattern Recognition : 28th International Conference, ICPR 2026, Lyon, France, August 17–22, 2026, Proceedings, Part XIV - 28th International Conference, ICPR 2026, Lyon, France, August 17–22, 2026, Proceedings, Part XIV
- series title
- Springer Lecture Notes in Computer Science
- volume
- 16825
- pages
- 16 pages
- publisher
- Springer
- conference name
- Pattern Recognition
- conference location
- Lyon, France
- conference dates
- 2026-08-17 - 2026-08-22
- external identifiers
-
- scopus:105047046136
- ISSN
- 0302-9743
- 1611-3349
- ISBN
- 978-3-032-31930-2
- 978-3-032-31929-6
- DOI
- 10.1007/978-3-032-31930-2_28
- language
- English
- LU publication?
- yes
- id
- bf18201c-235e-4388-b8d2-4f472506958f
- date added to LUP
- 2026-09-01 14:08:16
- date last changed
- 2026-09-10 03:37:11
@inproceedings{bf18201c-235e-4388-b8d2-4f472506958f,
abstract = {{Localization of players is a fundamental step in sports analytics and automated broadcasting, yet high-performance models for 4K resolution inputs create a significant computational bottleneck for real-time processing. Recent decomposable methods, such as DTSPL-BEV, address this by processing image “tiles” but rely on static, single-frame optimization and unstable heuristic merging algorithms. In this work, we propose EAST-SPL, an Event-Aware Statistical Tiling framework that shifts the optimization paradigm from minimizing single-frame FLOPs to minimizing Expected Total FLOPs (ETF). By leveraging the temporal distribution of player locations, our objective function prioritizes granular tiling in high-probability zones while allowing larger tiles in background regions. To realize these statistical gains, we introduce a lightweight auxiliary Rejection Network sharing convolutional features with the primary backbone to efficiently bypass empty tiles. Furthermore, we employ a Genetic Algorithm to navigate the non-differentiable configuration space, eliminating the need for heuristic merging. Experiments on the SynLoc dataset demonstrate that EAST-SPL reduces the expected computational cost by over 84% compared to conventional single-frame baselines, while also achieving substantially faster optimization than exhaustive grid search. (Code is available at https://github.com/AbolfazlChM95/EAST-SPL.)}},
author = {{Chaman Motlagh, Abolfazl and Nilsson, Mikael}},
booktitle = {{Pattern Recognition : 28th International Conference, ICPR 2026, Lyon, France, August 17–22, 2026, Proceedings, Part XIV}},
isbn = {{978-3-032-31930-2}},
issn = {{0302-9743}},
keywords = {{Object detection; Early rejection network; Decomposable Methods; Efficient ML; Genetic Algorithm}},
language = {{eng}},
month = {{08}},
pages = {{413--428}},
publisher = {{Springer}},
series = {{Springer Lecture Notes in Computer Science}},
title = {{EAST-SPL: Event-Aware Statistical Tiling for Decomposable Soccer Player Localization with an Auxiliary Rejection Network}},
url = {{http://dx.doi.org/10.1007/978-3-032-31930-2_28}},
doi = {{10.1007/978-3-032-31930-2_28}},
volume = {{16825}},
year = {{2026}},
}