In this talk, I will present a single-stage, uncertainty-aware approach
for jersey number recognition in soccer. My method uses
digit-compositional classifiers that exploit the structural
relationships between digits, integrated with a Dirichlet-based
uncertainty modeling framework. I also introduce a tracklet aggregation
strategy that combines frame-level predictions using confidence-based
filtering. Unlike traditional multi-stage pipelines that rely heavily on
explicit jersey detection, I reframe the problem from the perspective of
uncertainty quantification. This shift allows for more nuanced and
robust predictions, especially in cases of partial visibility or
occlusion. My unified architecture processes player crops directly,
simplifying the pipeline while enhancing reliability. Through extensive
experiments on the SoccerNet and Copa America (CA, my dataset)
benchmarks, I show that digit-compositional approaches consistently
outperform independent classifiers. Moreover, Dirichlet-based
uncertainty modeling yields better-calibrated confidence estimates
across varying visibility conditions. On the SoccerNet Challenge
benchmark, my method achieves strong results, reaching 85.62%
tracklet-level accuracy.