Jersey Number Tracking Solution

Our solution submitted on SoccerNet 2023

The SoccerNet Jersey Number Recognition Challenge 2023 required identifying soccer players' jersey numbers from short video tracklets, often spanning several hundred frames. With only one week to work on the challenge, there was limited time for model training and data collection, making it a particularly difficult task.

Example of Data Given For this Challenge

⚠️ Challenges:

  • Joining late left us with only one week, limiting time for model training and data collection.
  • Low-resolution, blurry images made it difficult to detect and recognize jersey numbers accurately.
  • Jersey numbers were visible briefly in a few frames, complicating consistent recognition.
  • Multiple players in tracklets created challenges in distinguishing and correctly tracking the target player.

💡 Solution:

Given the time limitations, we believed that a heuristic-based approachwould make the most sense, as it could offer a more controlled and efficient solution under the circumstances. We designed a Heuristic-Based Tracklet Aggregation Pipeline tailored to handle these challenges effectively.

Illustration of Our Solution Pipeline

🛠️ Proposed Method:

  • Pre-Processing: Enhanced input resolution and filtered out blurry frames using an ESRGAN-based super-resolution model.
  • Localization: Applied pose estimation to detect body keypoints, helping accurately localize jersey number regions based on player orientation.
  • Recognition: Performed digit recognition on localized tracklet segments.
  • Decision: Aggregated recognition results using a Multi-Step Aggregation Thresholds mechanism that factors in image quality and prediction confidence to determine the final jersey number.

🎯 Result:

  • Achieved 75.18% accuracy on the official challange set, outperforming several deep learning-based approaches on the leaderboard.
  • Demonstrated that a carefully designed heuristic pipeline can effectively compete with learning-based methods under strict data and quality constraints.

📌 Key Takeaway:

By leveraging super-resolution, pose-guided localization, and intelligent aggregation heuristics, we were able to show that an engineering-first, heuristic approach could achieve strong performance even when time and resources were limited.

📚 Related Articles