Research map / Vision-language and multimodal / Multimodal LLMs
SoccerMaster: A Vision Foundation Model for Soccer Understanding
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes SoccerMaster, a soccer-specific vision foundation model trained with supervised multi-task pretraining on spatial (athlete detection/identification, pitch registration) and semantic (event classification, vision-language alignment) tasks, using SoccerFactory, an automated annotation pipeline.
Problem
Soccer visual understanding is handled by fragmented task-specific expert models, and prior vision-language unification efforts neglect dense spatial objectives, so models describe what happens but not where or who.
Approach
A single video encoder (SigLIP 2-L initialized, with spatial then spatiotemporal attention blocks) is jointly pretrained with lightweight heads for detection, keypoint/line heatmaps, event classification, and contrastive alignment, using large-scale spatial labels automatically generated by the SoccerFactory pipeline from broadcast footage.
Key results
- SoccerMaster reaches 91.5 AP@50 and 49.5 mAP on athlete detection/identification versus 72.3 AP@50 and 32.0 mAP for the best baseline (SigLIP 2 with SoccerFactory data) in frozen-encoder comparison.
- Fine-tuned SoccerMaster achieves 75.8 FS on SN22-test-center and 56.2 FS on SN23-test at 512x512, beating PnLCalib at the same resolution (67.6 and 51.8).
- SoccerFactory annotations give a GS-HOTA of 64.1 on the SoccerNet-GSR test set, above challenge winner KIST-GSR (61.5).
Surprising findings
- Zero-shot SoccerMaster is below PnLCalib (960x540) on SN23-test calibration (44.8 vs 51.8 FS), so the advantage relies on fine-tuning or the resolution-matched baseline.
- In ablation, multi-task pretraining degrades athlete detection in the compact model (AP@50 88.3 to 82.0), which the authors attribute to limited capacity.
Limitations
- Jersey number recognition is weak, since the null class dominates training data.
- Compact-variant ablation shows athlete detection degrades under multi-task training, suggesting capacity limits.
- Main benchmarks often evaluated at 512x512 rather than the higher resolution PnLCalib uses, complicating comparison.
Experimental setup
Datasets: SoccerNet-GSR SoccerNet-v2 MatchTime SoccerReplay-1988 SoccerNet-tracking SN-Caption-test-align
Benchmarks: SoccerNet-GSR SoccerNet-v2 camera calibration (SN22, SN23) SoccerNet-tracking SN-Caption-test-align
Metrics: AP@50 mAP jersey number accuracy role accuracy keypoint and line accuracy/precision/recall event accuracy top-1 retrieval JaC CR FS
Baselines: SigLIP 2 DINOv3 MatchVision PnLCalib YOLOv8+OC-SORT YOLOv8+StrongSORT++ YOLOv8+PRTReID KIST-GSR
Builds on: SigLIP 2 TimeSformer UniSoccer PnLCalib MOTIP
Closest papers by meaning
- Towards Universal Soccer Video Understanding (CVPR 2025 · similarity 0.92)
- SportR: A Benchmark for Multimodal Large Language Model Reasoning in Sports (ICLR 2026 · similarity 0.88)
- SPORTU: A Comprehensive Sports Understanding Benchmark for Multimodal Large Language Models (ICLR 2025 · similarity 0.87)
- Event2Tracking: Reconstructing Multi-Agent Soccer Trajectories Using Long-Term Multimodal Context (AAAI 2025 · similarity 0.87)
- ViSTec: Video Modeling for Sports Technique Recognition and Tactical Analysis (AAAI 2024 · similarity 0.85)
- Learning Skill-Attributes for Transferable Assessment in Video (NeurIPS 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.