Beyond the Baseline: Enhanced SoccerNet Game State Reconstruction

This repository contains the code and methodology for our high-performance pipeline for the SoccerNet Game State Reconstruction (GSR) Challenge. Our approach systematically enhances the official baseline model, addressing key bottlenecks to achieve a significant boost in accuracy. Our work secures the second-best performance among all participants of the 2024 Challenge, demonstrating a 120% improvement over the baseline on the challenge set.

Abstract

The SoccerNet Game State Reconstruction (GSR) challenge provides a crucial benchmark for analyzing player tracking and identification from broadcast soccer videos. While the official baseline model offers a solid foundation, its performance is hampered by key bottlenecks in complex scenarios. This work introduces a series of targeted enhancements to create a new, high-performance pipeline. We systematically address the primary weaknesses of the baseline by integrating four key improvements: (1) a fine-tuned YOLOv8 model for robust athlete detection, (2) a fine-tuned CLIP model for accurate jersey number recognition, (3) a novel team clustering method using SIGLIP embeddings, and (4) the "No Bells, Just Whistles" method for precise pitch localization. Evaluated on the SoccerNet-GSR dataset, our proposed pipeline demonstrates a remarkable performance increase, achieving a GS-HOTA score of 51.52 on the challenge set—a 120% improvement over the baseline.

---

🚀 Key Features & Enhancements

Our pipeline replaces critical components of the SoccerNet-GSR baseline to achieve near state-of-the-art results.

1. ⚽ Robust Athlete Detection:

   • Replaced the standard YOLOv8x with a model fine-tuned on the SoccerNet v3 H250 dataset.
   • This significantly reduces false positives (e.g., detecting fans) and false negatives (e.g., missing occluded players).

2. 🔢 Accurate Jersey Number Recognition:

   • Replaced the baseline's MMOCR with a fine-tuned CLIP (ViT-L-14) model.
   • Overcomes failures of traditional OCR on low-resolution text and varied fonts by leveraging powerful vision-language representations.
   • Includes a data cleaning step to correct label noise in the original dataset.

3. 👕 Superior Team Clustering:

   • Replaced K-means clustering on ReID embeddings with a more robust method using pre-trained SIGLIP embeddings.
   • Accurately separates teams even when uniforms are visually similar.

4. 🗺️ Precise Pitch Localization:

   • Replaced the baseline's TVCalib with the state-of-the-art "No Bells, Just Whistles" (NBJW) method.
   • Achieves highly accurate camera calibration and player projection, especially in challenging center-pitch camera views.

📊 Performance

Our enhanced pipeline demonstrates a substantial improvement over the baseline across all official SoccerNet-GSR datasets. The core evaluation metric is GS-HOTA, which measures both localization and identification accuracy.

Challenge Set Results

Our method achieved the 2nd rank on the official challenge leaderboard.

Rank	Participant team	GS-HOTA(↑)	GS-DetA(↑)	GS-AssA(↑)
1	Constructor tech	63.81	49.52	82.23
2	Ours	51.52	37.35	71.07
3	UPCxMobius	43.15	30.46	61.16
12	Baseline	23.36	9.80	55.69

Performance Summary

Dataset	Baseline GS-HOTA	Our GS-HOTA	Improvement
Validation	18.05	37.10	+105%
Test	29.53	50.11	+69%
Challenge	23.36	51.52	+120%

---