GTATrack: Winner Solution to SoccerTrack 2025 with Deep-EIoU and Global Tracklet Association

TL;DR

GTATrack addresses the challenging task of multi-object tracking in fisheye soccer videos by coupling a motion-agnostic online association (Deep-EIoU) with a global trajectory refinement module (GTA-Link), augmented by a pseudo-labeling strategy to boost detector recall on small targets. The approach combines YOLOv11x-based detection with OSNet-based ReID to form robust short-term associations, followed by offline cluster-based linkages that merge fragments and resolve identity switches at the trajectory level. On SoccerTrack Challenge 2025, GTATrack achieves a leading HOTA of $0.60$ and a markedly reduced false-positive count of $982$, demonstrating strong performance under extreme distortion and uniform team jerseys. Overall, the paper demonstrates that hierarchical local-global reasoning, supported by pseudo-labeling and a robust ReID backbone, yields state-of-the-art accuracy for fisheye-based sports MOT with practical, scalable complexity.

Abstract

Multi-object tracking (MOT) in sports is highly challenging due to irregular player motion, uniform appearances, and frequent occlusions. These difficulties are further exacerbated by the geometric distortion and extreme scale variation introduced by static fisheye cameras. In this work, we present GTATrack, a hierarchical tracking framework that win first place in the SoccerTrack Challenge 2025. GTATrack integrates two core components: Deep Expansion IoU (Deep-EIoU) for motion-agnostic online association and Global Tracklet Association (GTA) for trajectory-level refinement. This two-stage design enables both robust short-term matching and long-term identity consistency. Additionally, a pseudo-labeling strategy is used to boost detector recall on small and distorted targets. The synergy between local association and global reasoning effectively addresses identity switches, occlusions, and tracking fragmentation. Our method achieved a winning HOTA score of 0.60 and significantly reduced false positives to 982, demonstrating state-of-the-art accuracy in fisheye-based soccer tracking. Our code is available at https://github.com/ron941/GTATrack-STC2025.

Figures (6)

• Figure 1: An overview of the complex tracking environment in a fisheye camera view. This frame illustrates several key challenges simultaneously: extreme scale variation between distant and nearby players, unpredictable spatial distribution reflecting irregular motion, and noticeable geometric distortion in peripheral areas. A robust tracker must effectively handle all these issues to maintain high accuracy.
• Figure 2: Illustration of tracking challenges in non-central regions of a fisheye camera view. The goalmouth area, a critical zone for gameplay, suffers from severe geometric distortion. Players in this region appear small and are often densely clustered, leading to frequent and prolonged occlusions.
• Figure 3: Overview of our GTATrack framework. (1) Object detection with a detector (e.g., YOLOv11x yolov11). (2) Appearance feature extraction using a ReID model (e.g., OSNet osnet-reid). (3) Online tracking via Deep-EIoU deepeiou to form initial tracklets. (4) Offline refinement with GTA-Link gta to merge fragments and correct identity switches.
• Figure 4: Deep-EIoU deepeiou adopts a multi-stage matching strategy that combines Expansion IoU and ReID features to prioritize high-quality associations. An iterative scale-up mechanism expands the spatial search area, improving robustness to occlusion and irregular motion.
• Figure 5: GTA-Link gta includes a Splitter for identity separation and a Connector for trajectory merging via spatio-temporal and appearance cues.