Broadcast2pitch++: depth-aware game state reconstruction from unconstrained soccer videos
TL;DR
Broadcast2Pitch++ is presented, an extended framework for robust game-state reconstruction from broadcast soccer footage that combines appearance similarity, motion consistency, and vision-language-derived identity attributes within a unified soft merge formulation.
Abstract
Game State Reconstruction (GSR) aims to recover the spatial positions and semantic identities of all players from broadcast soccer videos, requiring robust localization, tracking, and identity association under unconstrained camera motion and severe visual ambiguity. In this work, we present Broadcast2Pitch++, an extended framework for robust game-state reconstruction from broadcast soccer footage. We introduce two major extensions to the original framework for enhanced temporal identity consistency. First, we propose a depth-aware tracking formulation that incorporates monocular depth cues into the DeepEIoU association process. Second, we propose an improved tracklet refinement strategy that combines appearance similarity, motion consistency, and vision-language-derived identity attributes within a unified soft merge formulation. Comprehensive experiments on the SoccerNet-Tracking and SoccerNet-GSR benchmarks demonstrate that the proposed extensions consistently improve temporal association consistency in challenging broadcast soccer scenarios.