What Do Players Really See on the Pitch?
Using Pose Data, Field-of-View Models, and Pitch Control to Measure Visual Awareness in Soccer.
The following summary reviews the research paper titled “Wide Open Gazes: Quantifying Visual Exploratory Behavior in Soccer with Pose Enhanced Positional Data” by Joris Bekkers. All data, figures, and analysis presented here are drawn from their original work; I do not claim any authorship or ownership of the content. This summary has been written to provide a concise and technically informed synthesis of the paper’s findings, methodologies, and implications, while maintaining fidelity to the authors’ intellectual contributions.
1. Introduction
The paper situates itself within the broader evolution of soccer analytics, where event data has enabled action valuation, role identification, and passing structure analysis, while tracking data has unlocked spatial modeling, reinforcement learning approaches, and off-ball evaluation. This progression ultimately led to the development of pitch control and pitch value models, which quantify “the probability of each team controlling different areas of the pitch” and assign value to those areas.
Despite these advances, visual perception remains underexplored in real match contexts. Previous research has largely relied on laboratory-based eye-tracking studies, which, while insightful, fail to capture the “complex, dynamic demands of real soccer performance”. As a result, the translation to real-game measurement has been limited.
The dominant practical proxy has been Visual Exploratory Actions (VEAs), defined as rapid head movements exceeding 125°/s. However, the paper highlights multiple limitations: positional bias toward midfielders, annotation difficulty, binary measurement constraints, and weak predictive validity. Crucially, VEA-based metrics often fail to show statistically significant relationships with performance and may not differentiate elite from sub-elite players.
To address these issues, the study introduces a “stochastic vision layer” that models visual perception continuously using pose-enhanced tracking data. By combining head and shoulder orientation with spatial models, the approach represents visual attention as a probability distribution across the pitch rather than discrete scanning events.
This vision layer is integrated with pitch control and pitch value frameworks, allowing the author to link what a player sees to what they subsequently do. The key idea is that even aggregated visual metrics, such as the percentage of defensive space observed while awaiting a pass, can predict future gains in pitch value.
The approach offers several advantages over traditional methods. It is independent of player position, eliminates manual annotation, and produces continuous outputs that integrate naturally with existing analytical frameworks. Importantly, the paper argues that traditional VEA counts “lack predictive power for subsequent on-ball performance outcomes,” reinforcing the need for a more nuanced representation of perception.
1.1 Research
The methodology begins by modeling an individual player’s vision layer, composed of two elements: a probabilistic field of view and a probabilistic occlusion map. These are combined with pitch control and pitch value surfaces through element-wise multiplication.
A key extension is the introduction of “imminent pitch control,” which represents the space a player can reach in a very short time horizon. This allows the model to focus on actionable space rather than theoretical long-term control.
The framework is applied specifically to the “awaiting phase,” defined as the period when a player anticipates receiving a pass, and the subsequent on-ball phase. The central hypothesis is that visual exploration during this waiting period influences the quality of decisions once the ball is received.
To evaluate this, the author constructs binary classification models using XGBoost to predict whether a player’s pitch value increases or decreases after receiving the ball. The setup directly links perception (vision layer) to outcome (change in pitch value).
1.2 Validation Dataset
The dataset consists of 32 Copa América 2024 matches, combining broadcast tracking data with pose estimation and synchronized event data. The tracking data includes player positions as well as head, shoulder, and hip orientations, enabling the construction of detailed vision models.
Event data is aligned with tracking using synchronization techniques, and ball receptions are defined as local minima in player-ball distance. The dataset contains approximately 2.3 million frames and over 60,000 events, filtered to focus on uncontested open-play sequences.
From this, around 14,000 “awaiting moments” are extracted, each followed by an on-ball action. Missing pose angles are interpolated using Euler-based methods to ensure continuity in orientation data.
2. Methodology
2.1 Vision
The vision model constructs a two-dimensional probabilistic representation of what a player can see. It consists of two components: a field of view model and an occlusion model, combined into a full vision map over a 105×68 grid.
A key assumption is that head orientation approximates gaze direction, supported by evidence that “approximately 92% of gaze shifts are in the same direction as head movement”. The model accounts for uncertainty, recognizing that perception degrades with distance, peripheral angle, and movement speed.
2.1.1 Field of View
The field of view is first defined as a binary region spanning 120°, representing human binocular vision. This is then transformed into a probabilistic surface that reflects the likelihood of accurately perceiving different locations.
The probabilistic field of view is given by:
where radial and angular Gaussian components control how vision decays with distance and angle. Both components are influenced by player speed, based on the assumption that higher speed reduces spatial awareness and narrows attention.
This produces a smooth, continuous vision map where central areas are perceived more clearly than peripheral or distant ones.
2.1.2 Occlusions
The occlusion model accounts for other players blocking vision. Each opponent or teammate generates a probabilistic shadow, whose size depends on distance and body orientation.
Occlusion is modeled through a probabilistic ray extending from the observing player through the obstructing player.
The width of this ray depends on the apparent angular size of the obstructing player, which is derived from their shoulder orientation and physical dimensions.
The combined occlusion map is constructed as the product of all individual occlusion effects.
This ensures that multiple players jointly reduce visibility in overlapping regions.
2.1.3 Complete Vision Map of Player i
The final vision map is obtained by combining field of view and occlusion.
This map represents, for every location on the pitch, the probability that the player can visually perceive it at a given moment. Importantly, it is continuous, context-dependent, and dynamically updated.
2.2 Pitch Control
To link perception with actionable space, the model integrates pitch control and pitch value concepts from prior work.
2.2.1 Imminent Pitch Control
The standard pitch control model estimates the probability that a player can reach a location first. The author modifies this by introducing a scaling parameter to reduce the influence radius, creating “imminent pitch control.”
This represents the space a player can realistically influence in the short term, aligning better with the temporal constraints of decision-making.
Different versions are computed for attacking players, defending players, and all players except the focal player, allowing flexible analysis of spatial context.
2.2.2 Pitch Value
Pitch value is modeled using a neural network that predicts the strategic importance of each location given the ball position. This is based on the idea that defensive positioning implicitly reveals valuable attacking spaces.
The final pitch value surface is normalized by distance to goal, reflecting that areas closer to goal are inherently more valuable. The resulting surface ranges between 0 and 1 and varies dynamically with ball location.
3. Validation
The validation assesses whether visual perception during the awaiting phase predicts subsequent improvements in pitch value.
The dataset includes over 170,000 frames aligned with 14,000 awaiting moments. The core task is binary classification: whether pitch value increases or decreases after the on-ball phase.
3.1 Labels
The label is defined using a ratio of pitch value before and after the on-ball phase.
Samples with ambiguous changes are excluded to reduce noise, ensuring clear positive and negative cases.
3.2 Ablation Models
Four models are compared:
Baseline,
VEA-based model
Regular feature model
Vision-enhanced model.
The vision model includes features derived from the interaction between vision maps and pitch control surfaces, such as the proportion of observed attacking or defensive space and their temporal averages.
3.3 Validation Results
The results show a clear improvement when vision features are included. The AUC increases from 0.744 in the regular model to 0.788 in the vision model, demonstrating the added predictive value.
In contrast, adding traditional VEA features does not improve performance, reinforcing their limited usefulness as a metric.
3.4 Impactful Features
Feature importance analysis reveals that the most predictive variables relate to the quality of observed space rather than the quantity.
Observing larger amounts of occupied space, particularly defensive space, during the awaiting phase is associated with better outcomes. This suggests that effective players identify constraints and opportunities before receiving the ball.
Conversely, simply observing more of the pitch does not improve performance, highlighting that perception quality matters more than coverage.
4. Conclusion
The paper presents a novel framework for modeling visual perception in soccer using pose-enhanced tracking data. By integrating probabilistic vision maps with pitch control and pitch value models, it provides a continuous and context-aware measure of visual exploratory behavior.
The findings show that visual behavior during the awaiting phase has predictive power for subsequent on-ball success. Players who observe more relevant space, particularly defensively controlled areas, tend to achieve greater improvements in pitch value.
Importantly, traditional VEA metrics are shown to have “zero predictive value” in this context, emphasizing the limitations of binary scanning measures.
The approach is scalable, position-independent, and compatible with existing analytical frameworks, with potential applications beyond soccer as pose data becomes more widely available.
Presented at the MIT Sloan Sports Analytics Conference
This research was also presented at the MIT SSAC 2026, where the author broke down the models, visuals, and implications in more depth.
If you want to see the framework explained directly by the person who built it, the full talk is worth your time: https://www.youtube.com/watch?v=z8er_Z4WZS0
References
Bekkers, J. (2026). Wide Open Gazes: Quantifying Visual Exploratory Behavior in Soccer with Pose Enhanced Positional Data. arXiv preprint arXiv:2602.18519.
https://arxiv.org/abs/2602.18519
Learn More
Here you have a direct link to get the free xG Mini-Guide, suited for those who want to fully understand the model in 5 minutes.
