SoccerTrack v2: Full-Pitch Tracking Dataset

• SoccerTrack v2 is a full-pitch soccer tracking dataset capturing 10 complete university matches in panoramic 4K with synchronized positional and event annotations.
• It supports research in multi-object tracking, game state reconstruction, and ball action spotting by providing persistent player IDs, role metadata, and BAS labels for 12 action classes.
• The dataset enables diverse tracking paradigms—from fisheye camera challenges to graph-based multi-camera analysis—benchmarking performance with metrics like HOTA, DetA, and AssA.

Searching arXiv for SoccerTrack v2 and related tracking papers to ground the article in current literature. SoccerTrack v2 is a public dataset for soccer analytics that is purpose-built to advance multi-object tracking (MOT), game state reconstruction (GSR), and ball action spotting (BAS) in full-pitch settings. It is defined by 10 full-length university-level matches recorded in panoramic 4K with BePro systems, with synchronized positional and event annotations designed for long-sequence tracking, minimap-style reconstruction, and event-centric analysis. Relative to broadcast-view corpora, its central distinguishing property is full-pitch coverage with near-complete player visibility, together with per-frame pitch-coordinate annotations, persistent identities, role and team metadata, and BAS timestamps for 12 action classes (Scott et al., 3 Aug 2025).

1. Dataset definition and scope

SoccerTrack v2 targets three tasks in a unified setting: MOT, GSR, and BAS. The dataset consists of 10 full-length university-level matches, approximately 900 minutes total, recorded as panoramic 4K MP4 videos using BePro systems. Two matches used BePro Cerberus, while eight used BePro’s standard 3‑camera panoramic stitching system. The recordings are fixed panoramic full-pitch setups intended to preserve end-to-end pitch coverage and all-player visibility, including day and night conditions, multiple locations, and different weather conditions (Scott et al., 3 Aug 2025).

The dataset is positioned as a response to limitations of prior broadcast-oriented resources. Unlike SoccerNet and SportsMOT, which rely on broadcast views and short clips, SoccerTrack v2 provides full-pitch panoramic 4K coverage for complete matches. Unlike SoccerTrack-v1 and TeamTrack, it expands to multiple matches and includes jersey numbers and role labels. The inclusion of BAS labels for 12 action classes further distinguishes it as a dataset for unified tracking-plus-event analysis rather than tracking alone (Scott et al., 3 Aug 2025).

A plausible implication is that SoccerTrack v2 is structured for research questions that require persistent team-level context over long horizons, such as formation tracking, compactness analysis, or pre-event spatial conditioning, which are difficult to study in short broadcast clips.

2. Recording setup, synchronization, and annotation schema

The recording setup centers on BePro panoramic capture. The public dataset emphasizes panoramic full-pitch videos and synchronized positional and event annotations. Player positional data and event logs from BePro were synchronized to the video timeline using global timestamps. The report states that the dataset uses BePro-provided player positions in pitch coordinates, expressed in meters, while specific camera intrinsics, extrinsics, or homography files are not explicitly stated (Scott et al., 3 Aug 2025).

The GSR annotation schema is per frame. Each visible player, goalkeeper, and referee is assigned 2D pitch coordinates in meters, a unique track ID persistent throughout the match, a role label, a team side label, and a jersey number when visible. The allowed role values are player, goalkeeper, referee, and other; team side is left, right, or null; jersey number is an integer from 0–99 if visible and null otherwise. The coordinate system is directly in 2D pitch coordinates and is described as consistent per match, following a SoccerNet-like positional format (Scott et al., 3 Aug 2025).

The BAS annotation schema consists of a global timestamp aligned to the video and one action class from 12 categories: Pass, Drive, Header, High Pass, Out, Cross, Throw In, Shot, Ball Player Block, Player Successful Tackle, Free Kick, and Goal. The report does not state that per-frame ball positions or ball possession labels are provided; BAS is event-based rather than trajectory-based (Scott et al., 3 Aug 2025).

The annotation process combines synchronized BePro logs with manual checks and corrections on events. Full bounding boxes for all frames were initially planned but considered too labor-intensive, at approximately 5000 hours for about 1.62 million frames across 10 matches. Instead, a curated subset with bounding boxes and player IDs is designated for the SoccerTrack Challenge (MMSports 2025) (Scott et al., 3 Aug 2025).

Annotation component	Granularity	Contents
GSR	Per frame	2D pitch coordinates, persistent IDs, roles, team side, jersey numbers when visible
BAS	Timestamped	12 action classes aligned to video time
Bounding boxes	Curated subset	Planned for SoccerTrack Challenge subset rather than all frames

This schema makes SoccerTrack v2 unusual in that it provides game-state labels directly in pitch coordinates. For many GSR applications, this eliminates the need to estimate image-to-pitch mappings before tactical analysis.

3. Evaluation tasks, metrics, and geometric interpretation

SoccerTrack v2 is positioned as a benchmark for MOT, GSR, and BAS, although the technical report does not provide numerical baseline results. Instead, it specifies the typical evaluation protocols used for each task. For MOT, the commonly cited metrics are MOTA, IDF1, and HOTA, with HOTA decomposed into components such as Association Accuracy (AssA) and Localization Accuracy (LocA). For BAS, standard event spotting metrics include mAP computed over temporal IoU or tolerance windows and F1 score at fixed spotting tolerance. For GSR, position error in meters is measured using metrics such as MAE and RMSE between predicted and ground-truth 2D pitch coordinates (Scott et al., 3 Aug 2025).

The report explicitly gives the standard formulas:

$$\mathrm{MOTA} = 1 - \frac{\mathrm{FN} + \mathrm{FP} + \mathrm{IDsw}}{\mathrm{GT}}$$

$$\mathrm{IDF1} = \frac{2 \cdot \mathrm{IDP} \cdot \mathrm{IDR}}{\mathrm{IDP} + \mathrm{IDR}}$$

$$\mathrm{MAE} = \frac{1}{N} \sum_{i=1}^{N} \| \hat{\mathbf{x}}_i - \mathbf{x}_i \|_2, \qquad \mathrm{RMSE} = \sqrt{\frac{1}{N} \sum_{i=1}^{N} \| \hat{\mathbf{x}}_i - \mathbf{x}_i \|_2^2}$$

Because GSR positions are already provided in pitch coordinates, many applications can operate without explicit calibration. However, the report also frames standard geometry for reconstructing game state from image detections. For planar pitch points, it gives a homography relation

$$\begin{bmatrix} x_p \ y_p \ 1 \end{bmatrix} \sim \mathbf{H} \begin{bmatrix} u \ v \ 1 \end{bmatrix},$$

and for calibrated cameras it gives the projective camera model

$$s \begin{bmatrix} u \ v \ 1 \end{bmatrix} = \mathbf{K} [\mathbf{R}\ |\ \mathbf{t}] \begin{bmatrix} X \ Y \ Z \ 1 \end{bmatrix}.$$

It also notes that if raw multi-view feeds are used, standard linear triangulation can estimate 3D points before projection to the pitch plane (Scott et al., 3 Aug 2025).

This suggests that SoccerTrack v2 supports two distinct methodological regimes: direct use of provided pitch-space state labels for downstream analytics, and image-based reconstruction pipelines that predict detections or tracks and then map them into pitch coordinates.

4. Challenge setting and benchmarked tracking on SoccerTrack v2

In the challenge context, SoccerTrack v2 is used for single-camera tracking in full-pitch soccer videos captured by static fisheye cameras. The winning method, GTATrack, frames the problem as frame-wise detection and identity assignment for soccer players over entire matches, without modeling the ball. The paper emphasizes strong barrel or pincushion distortions, non-uniform spatial resolution, extreme scale variation, uniform appearances, irregular motion, and frequent occlusions as the defining technical difficulties of this benchmark configuration (Jian et al., 31 Jan 2026).

The challenge paper reports that its experiments use six full-pitch fisheye videos at 4096×1080 resolution with 22 unique player identities, with four videos used for training and two for validation and ablation. Leaderboard ranking is by HOTA on the official test split, and the paper reports HOTA together with DetA, AssA, LocA, FP, FN, and IDSW. The metric definitions are given explicitly:

$HOTA_\alpha = \sqrt{DetA_\alpha \cdot AssA_\alpha}$

with averaging over IoU thresholds for the final HOTA, DetA, AssA, and LocA values (Jian et al., 31 Jan 2026).

GTATrack is a hierarchical pipeline with two stages: Deep Expansion IoU (Deep-EIoU) for motion-agnostic online association and Global Tracklet Association (GTA-Link) for offline refinement. The detector is YOLOv11x, trained with multi-scale and mosaic augmentation and strengthened by pseudo-labeling on unlabeled frames. Online association combines spatial overlap through Expansion IoU with cosine-based OSNet appearance distance:

$$C_{ij} = \lambda_e \cdot [1 - EIoU(b_i, b_{\tau_j})] + \lambda_a \cdot d_{\text{app}}(i, j),$$

and matches are solved with the Hungarian algorithm. Offline refinement applies constrained hierarchical clustering over tracklets, with Splitter and Connector components to separate mixed-identity trajectories and merge fragments (Jian et al., 31 Jan 2026).

The official challenge result reported for GTATrack is first place with HOTA $= 0.60$, IDSW $= 331.5$, LocA $= 0.84$, DetA $$\mathrm{IDF1} = \frac{2 \cdot \mathrm{IDP} \cdot \mathrm{IDR}}{\mathrm{IDP} + \mathrm{IDR}}$$0, AssA $$\mathrm{IDF1} = \frac{2 \cdot \mathrm{IDP} \cdot \mathrm{IDR}}{\mathrm{IDP} + \mathrm{IDR}}$$1, FN $$\mathrm{IDF1} = \frac{2 \cdot \mathrm{IDP} \cdot \mathrm{IDR}}{\mathrm{IDP} + \mathrm{IDR}}$$2, and FP $$\mathrm{IDF1} = \frac{2 \cdot \mathrm{IDP} \cdot \mathrm{IDR}}{\mathrm{IDP} + \mathrm{IDR}}$$3. Ablations show that pseudo-labeling, Deep-EIoU, and GTA-Link each contribute materially, with pseudo-labeling alone improving HOTA from 0.380 to 0.491 and reducing FP and FN substantially (Jian et al., 31 Jan 2026).

A plausible implication is that the challenge subset operationalizes SoccerTrack v2 less as a generic full-pitch analytics corpus and more as a stress test for long-horizon MOT under fisheye distortion, where detection recall on tiny peripheral players and trajectory repair under severe occlusion become the main determinants of HOTA.

5. Tracking paradigms relevant to SoccerTrack v2

A broadcast-oriented but directly transferable tracking paradigm is described by the MOT4MOT technical report for the SoccerNet 2023 Tracking Challenge. Although that paper does not mention SoccerTrack or SoccerTrack v2 explicitly, its task structure maps directly to SoccerTrack v2-style player-and-ball tracking benchmarks. Its pipeline combines a fine-tuned YOLOv8l detector with DeepOC-SORT++ or StrongSORT++, followed by offline post-processing using Gaussian-process smoothing and interpolation (GSI), appearance-free linking (AFLink), and a soccer-specific appearance-based merging heuristic that re-links tracks terminating far from image boundaries (Shitrit et al., 2023).

For players, the strongest reported configuration is DeepOC-SORT++, using a constant-velocity Kalman filter, camera motion compensation, a fine-tuned OSNet-ain appearance model, cosine-similarity association, and offline repair of fragmented tracks. The standard Kalman prediction and update are given as

$$\mathrm{IDF1} = \frac{2 \cdot \mathrm{IDP} \cdot \mathrm{IDR}}{\mathrm{IDP} + \mathrm{IDR}}$$4

and

$$\mathrm{IDF1} = \frac{2 \cdot \mathrm{IDP} \cdot \mathrm{IDR}}{\mathrm{IDP} + \mathrm{IDR}}$$5

$$\mathrm{IDF1} = \frac{2 \cdot \mathrm{IDP} \cdot \mathrm{IDR}}{\mathrm{IDP} + \mathrm{IDR}}$$6

Interpolation across a gap $$\mathrm{IDF1} = \frac{2 \cdot \mathrm{IDP} \cdot \mathrm{IDR}}{\mathrm{IDP} + \mathrm{IDR}}$$7 is expressed as

$$\mathrm{IDF1} = \frac{2 \cdot \mathrm{IDP} \cdot \mathrm{IDR}}{\mathrm{IDP} + \mathrm{IDR}}$$8

For the ball, the method departs from MOT and treats tracking as single-object detection per frame, using a low-threshold YOLOv8l detector, a 3rd-order polynomial fit over a 51-frame window, proximity filtering within 100 pixels, and linear interpolation for missing detections (Shitrit et al., 2023).

The final reported performance on the SoccerNet challenge is HOTA $$\mathrm{IDF1} = \frac{2 \cdot \mathrm{IDP} \cdot \mathrm{IDR}}{\mathrm{IDP} + \mathrm{IDR}}$$9, DetA $$\mathrm{MAE} = \frac{1}{N} \sum_{i=1}^{N} \| \hat{\mathbf{x}}_i - \mathbf{x}_i \|_2, \qquad \mathrm{RMSE} = \sqrt{\frac{1}{N} \sum_{i=1}^{N} \| \hat{\mathbf{x}}_i - \mathbf{x}_i \|_2^2}$$0, AssA $$\mathrm{MAE} = \frac{1}{N} \sum_{i=1}^{N} \| \hat{\mathbf{x}}_i - \mathbf{x}_i \|_2, \qquad \mathrm{RMSE} = \sqrt{\frac{1}{N} \sum_{i=1}^{N} \| \hat{\mathbf{x}}_i - \mathbf{x}_i \|_2^2}$$1, yielding third place. The report also states that using ground-truth boxes instead of detections raises DeepOC-SORT++ HOTA from 66.00 to 87.85, indicating that detection quality remains the dominant bottleneck (Shitrit et al., 2023).

For SoccerTrack v2, the report explicitly states that this pipeline is directly applicable with minimal adaptation: fine-tune YOLOv8l and OSNet on SoccerTrack v2 annotations and identity splits, recalibrate thresholds to the dataset’s resolution and frame rate, retain boundary-aware appearance-based merging, and evaluate using HOTA, DetA, AssA, with optional addition of IDF1 and MOTA. This suggests that SoccerTrack v2 supports both fisheye-native trackers such as GTATrack and more conventional tracking-by-detection systems, provided they are retuned to full-pitch panoramic imagery.

6. Relation to multi-camera and pitch-space tracking research

SoccerTrack v2’s GSR formulation also aligns with a different tracking tradition: direct pitch-space tracking from synchronized multi-camera observations. A representative example is the graph-based multi-camera soccer player tracker of Vats et al., which operates on raw multi-camera detection heat maps projected into a global bird’s-eye coordinate frame rather than on per-camera bounding boxes and appearance cues. The method assumes 4–6 fixed calibrated cameras around the pitch and targets long-shot views where players are difficult to distinguish visually, so it emphasizes individual dynamics and local player interactions instead of appearance-based re-identification (Komorowski et al., 2022).

Its representation stacks projected per-camera heat maps into a multi-channel ground-plane tensor. For each active player track, a cropped region around the previous pitch position is encoded by a detection encoder $$\mathrm{MAE} = \frac{1}{N} \sum_{i=1}^{N} \| \hat{\mathbf{x}}_i - \mathbf{x}_i \|_2, \qquad \mathrm{RMSE} = \sqrt{\frac{1}{N} \sum_{i=1}^{N} \| \hat{\mathbf{x}}_i - \mathbf{x}_i \|_2^2}$$2, motion history is encoded by an LSTM $$\mathrm{MAE} = \frac{1}{N} \sum_{i=1}^{N} \| \hat{\mathbf{x}}_i - \mathbf{x}_i \|_2, \qquad \mathrm{RMSE} = \sqrt{\frac{1}{N} \sum_{i=1}^{N} \| \hat{\mathbf{x}}_i - \mathbf{x}_i \|_2^2}$$3, and neighborhood interactions are modeled by a graph neural network with edges between players within 3 meters. Message passing is given by

$$\mathrm{MAE} = \frac{1}{N} \sum_{i=1}^{N} \| \hat{\mathbf{x}}_i - \mathbf{x}_i \|_2, \qquad \mathrm{RMSE} = \sqrt{\frac{1}{N} \sum_{i=1}^{N} \| \hat{\mathbf{x}}_i - \mathbf{x}_i \|_2^2}$$4

$$\mathrm{MAE} = \frac{1}{N} \sum_{i=1}^{N} \| \hat{\mathbf{x}}_i - \mathbf{x}_i \|_2, \qquad \mathrm{RMSE} = \sqrt{\frac{1}{N} \sum_{i=1}^{N} \| \hat{\mathbf{x}}_i - \mathbf{x}_i \|_2^2}$$5

with edge attributes

$$\mathrm{MAE} = \frac{1}{N} \sum_{i=1}^{N} \| \hat{\mathbf{x}}_i - \mathbf{x}_i \|_2, \qquad \mathrm{RMSE} = \sqrt{\frac{1}{N} \sum_{i=1}^{N} \| \hat{\mathbf{x}}_i - \mathbf{x}_i \|_2^2}$$6

Final pitch positions are regressed as

$$\mathrm{MAE} = \frac{1}{N} \sum_{i=1}^{N} \| \hat{\mathbf{x}}_i - \mathbf{x}_i \|_2, \qquad \mathrm{RMSE} = \sqrt{\frac{1}{N} \sum_{i=1}^{N} \| \hat{\mathbf{x}}_i - \mathbf{x}_i \|_2^2}$$7

using the training loss

$$\mathrm{MAE} = \frac{1}{N} \sum_{i=1}^{N} \| \hat{\mathbf{x}}_i - \mathbf{x}_i \|_2, \qquad \mathrm{RMSE} = \sqrt{\frac{1}{N} \sum_{i=1}^{N} \| \hat{\mathbf{x}}_i - \mathbf{x}_i \|_2^2}$$8

On synthetic and real multi-camera datasets, the method improves MOTA and reduces ID switches relative to a particle-filter baseline, and ablations show substantial degradation when either the trajectory encoder or message passing is removed (Komorowski et al., 2022).

This line of work is not a direct benchmark on SoccerTrack v2, but it is highly relevant because SoccerTrack v2 provides per-frame pitch coordinates in meters. A plausible implication is that SoccerTrack v2 can serve not only as an image-based MOT dataset but also as a ground-truth substrate for learning pitch-space dynamics, calibrating image-to-pitch mappings, or evaluating graph-based interaction models.

7. Limitations, biases, availability, and research directions

Several limitations are explicit in the SoccerTrack v2 report. The matches are university-level rather than professional, so game pace and tactical sophistication may differ from elite environments. Jersey numbers can be absent or occluded and are then set to null. Detailed camera calibration files are not fully provided. Class imbalance in BAS is likely, but per-class counts are not reported. Full bounding boxes are not available for all frames because of annotation cost, and official train, validation, and test splits are not specified in the report itself (Scott et al., 3 Aug 2025).

The challenge and benchmark papers expose complementary technical bottlenecks. In fisheye single-camera tracking, severe distortion, tiny peripheral targets, and uniform appearances remain difficult even with pseudo-labeling and global tracklet refinement (Jian et al., 31 Jan 2026). In broadcast-style pipelines, detector quality, missed detections, heavy occlusions, overlapping players, and ball confusion with white shoes or crowd elements remain dominant failure modes (Shitrit et al., 2023). In multi-camera pitch-space tracking, calibration drift, synchronization errors, and ambiguity under dense occlusion remain limiting factors even when appearance is deemphasized (Komorowski et al., 2022).

The dataset is to be hosted on Hugging Face and GitHub, with download scripts and checksum verification. The technical report states that researchers must cite it when publishing, that ethical approvals were obtained, and that personal identities are anonymized through jersey numbers only. A curated subset with bounding boxes and player IDs is designated for the SoccerTrack Challenge (MMSports 2025), with evaluation scripts and splits to accompany that release (Scott et al., 3 Aug 2025).

Taken together, SoccerTrack v2 occupies a distinctive position in soccer vision research. It combines full-pitch panoramic video, per-frame pitch-coordinate state labels, persistent identity metadata, and BAS events over full matches, while also supporting challenge-style MOT evaluation under fisheye full-pitch imagery. This combination makes it simultaneously a dataset for long-horizon tactical reconstruction, a benchmark for difficult soccer MOT, and a substrate for integrating detection, association, geometry, and event understanding within a single soccer-specific experimental framework.