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

AI summary

Overview

• Research area: Multi-object tracking (MOT) in sports video, specifically soccer footage captured by static fisheye cameras.
• Technical level: Intermediate. The paper assumes familiarity with tracking-by-detection pipelines, HOTA-style metrics, and ReID embeddings, but it is a system-level engineering paper rather than a deep methodological one.
• Scope: This paper describes GTATrack, a two-stage local-plus-global tracking framework built from existing components (YOLOv11x, OSNet, Deep-EIoU, GTA-Link) that won first place in the SoccerTrack Challenge 2025.

What This Paper Is About

Tracking players in broadcast or fixed-camera soccer is hard because players move erratically, wear nearly identical uniforms, and constantly occlude one another. The SoccerTrack Challenge 2025 adds a static fisheye camera on top of these problems, producing severe geometric distortion, extreme scale variation, and low resolution for distant players. The goal of this work is to build a tracker that keeps player identities consistent across long sequences under all of these conditions simultaneously.

Key Contributions

1. GTATrack, a hierarchical local-global tracking framework that combines motion-agnostic online association with global trajectory refinement for stable tracking in fisheye soccer videos.
2. Deep-EIoU for robust local association, using iterative bounding box expansion and deep appearance features to handle irregular motion without relying on predictive motion models such as Kalman filters.
3. GTA-Link for global trajectory refinement, which resolves long-term identity switches by clustering fragmented tracklets using spatio-temporal and appearance-based reasoning.
4. A pseudo-labeling strategy for the detector, which boosts recall on small and distorted targets and is credited in the conclusion with reducing false positives by nearly 90%.

Main Findings

• First place in SoccerTrack Challenge 2025: GTATrack, entered as team "wccjs," ranked first with a HOTA of 0.60, ahead of mdk-tdu (0.59), SShota (0.59), takashun13 (0.59), and YMori22 (0.59).
• Very low false positives: The winning entry reported 982.00 FP, the second-best value on the leaderboard, achieved through the pseudo-label-enhanced detector.
• Ablation on the validation set: ByteTrack (baseline) reached HOTA 0.42 with 630 IDSW; Deep-EIoU reached HOTA 0.54 with 325.5 IDSW; Deep-EIoU + GTA-Link with finetuning reached HOTA 0.60 with 331.5 IDSW, 0.84 LocA, 0.76 DetA, and 0.47 AssA (up from 0.26 for ByteTrack).
• Note on reported numbers: The body text describes the Deep-EIoU gain as HOTA 0.52 to 0.55 and AssA 0.38 to 0.42, which differs from the 0.42 → 0.54 and 0.26 → 0.38 figures in the ablation table. The table values are treated here as the reported ablation results.
• Pseudo labels help substantially: Without pseudo labels, HOTA was 0.380 with 4913 FP and 40046 FN; with pseudo labels, HOTA was 0.491 with 494 FP and 16186 FN. The text narrative states a HOTA increase "to 0.511," which again differs from the table value of 0.491.
• Detector and ReID combination matters: Using a fixed Deep-EIoU backbone, YOLOv11x + OSNet scored HOTA 0.491; YOLOv11x + SOLIDER scored 0.474; SO-DETR + OSNet scored 0.405; SO-DETR + SOLIDER scored 0.357. The text describes the best configuration as HOTA 0.511 and the worst as 0.357.
• OSNet beat the Transformer-based ReID: OSNet outperformed the Transformer-based model across all detector pairings, with the authors attributing this to its lightweight design and multi-scale feature extraction being more resilient to occlusion and fast motion, whereas Transformer-based ReID showed instability under high motion.
• Deep-EIoU proximity threshold matters: Raising the threshold from 0.4 to 0.9 improved HOTA from 0.491 to 0.547, an 11.4% gain, with best results in the 0.8–0.9 range. Setting it to 1.0 removed spatial filtering entirely and caused mismatches between visually similar but spatially distant players.
• GTA-Link adds 3%–4% HOTA for identity consistency across parameter settings. The original tracklet count without GTA-Link was 53; under eps = 0.5 and min-samples = 7, the system split 53 tracklets into 69 fragments and merged them into 27 refined trajectories.
• The Splitter is scenario-dependent: The paper reports that its benefit may cause over-fragmentation in stable scenes and suggests an adaptive activation strategy could improve robustness.

Methodology in Plain English

The researchers built a pipeline with four stages rather than training one large end-to-end model.

First, a detector finds players in every frame. They considered two detectors, YOLOv11x and SO-DETR, and selected YOLOv11x because it handled small and distant players well while being more efficient. To improve recall, they used a semi-supervised pseudo-labeling step: the fine-tuned YOLOv11x generated predictions on unlabeled frames, high-confidence detections were kept as pseudo-labels, and those were added back into training.

Second, a person re-identification model turns each detected player crop into a normalized feature vector. They compared the CNN-based OSNet against the Transformer-based SOLIDER and chose OSNet. These vectors are compared with cosine distance to judge whether two detections are the same person.

Third, an online tracker called Deep-EIoU links detections into tracklets frame by frame. Instead of predicting where a player will be next using a motion model, it expands bounding boxes iteratively and combines spatial overlap with appearance similarity into a single cost, which is then solved as an assignment problem with the Hungarian algorithm. This avoids assuming that motion is smooth, which is wrong for athletes.

Fourth, an offline module called GTA-Link revisits the finished tracklets. It treats each tracklet as a node and merges fragments using hierarchical clustering on average pairwise appearance distance, subject to spatial and temporal constraints and a similarity threshold. A Tracklet Splitter can also separate multiple identities wrongly merged into one tracklet, though the authors focus on the Connector in this work.

Experiments used the official SoccerTrack Challenge 2025 dataset of fisheye soccer videos at 4096×1080 resolution with 22 unique player identities. Four videos were used for training and two for validation in ablations. Training ran on a single NVIDIA RTX 3090 GPU for 200 epochs using AdamW with a learning rate of 0.0001, multi-scale and mosaic augmentation, a detector input resolution of 1280 pixels on the longer side, and a batch size of 12. Evaluation used HOTA as the primary metric, alongside IDSW, AssA, LocA, and DetA.

Why This Matters

Impact on research: The paper shows that a carefully assembled pipeline of existing components can beat purpose-built alternatives in a specific hard domain, and it provides an ablation-driven account of which pieces matter. It also reinforces the value of motion-agnostic association over Kalman-filter-based prediction in scenarios with abrupt, non-linear movement, and it highlights fisheye distortion as an underexplored stress test for tracking and ReID models.

Real-world applications:

• Automated tactical and performance analytics for soccer teams, where consistent player identities are a prerequisite for computing distance covered, positioning, and formation statistics.
• Broadcast production tools such as automatic player highlighting, replay framing, and augmented graphics that need reliable identity tracking through occlusions.
• Sports officiating support and post-match review, where player trajectories and interaction histories must remain attributable to the correct individual.
• General surveillance and analytics in wide-angle or surround-view camera setups, where the same distortion, scale-variation, and identity-switch problems appear.

Industry relevance: The paper is a competition-winning system released as open code (https://github.com/ron941/GTATrack-STC2025), which makes it a practical starting point for commercial sports-analytics platforms. The finding that a compact CNN ReID model (OSNet) beat a Transformer-based one in this high-motion setting is also relevant to engineers balancing accuracy against latency and memory in deployment.

Future Directions

• Adaptive Splitter activation: The paper notes that the Tracklet Splitter's benefit is scenario-dependent and can over-fragment stable scenes, suggesting an adaptive activation strategy as a route to more robustness.
• Closing the gap between local and global reasoning: The hierarchical design currently separates online association and offline refinement; whether these could be unified into a single end-to-end trainable tracker is left open.
• Revisiting Transformer-based ReID: The observed instability of SOLIDER under high motion raises the question of whether geometry-invariant or motion-aware re-identification objectives could make attention-based models competitive in fisheye scenes.
• Generalization beyond this dataset: The evaluation is confined to the SoccerTrack Challenge 2025 fisheye dataset and a six-video statistics table; whether GTATrack transfers to other sports, other camera geometries, or larger multi-sport benchmarks is not reported.

Target Audience

This paper is most useful to computer vision engineers and researchers working on multi-object tracking, particularly those in sports analytics and video surveillance. It also suits practitioners who need a working, competition-validated reference implementation rather than a new theoretical contribution, and graduate students studying tracking-by-detection pipelines, re-identification in low-inter-class-variance settings, or distortion-aware perception. Readers primarily interested in training dynamics, loss design, or benchmark-wide generalization will find it is not that kind of paper.

Authors’ 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.