Measuring the Unmeasurable: Why Computer Vision May Finally Prove Momentum Is Real

Super Bowl LI wasn't close for three quarters. The Atlanta Falcons dismantled the Patriots with the confidence of a team that had shredded defenses all season. Up 28-3 midway through the third quarter, Atlanta's win probability stood at 99.8 percent. Even Vegas thought it was over.

Then something broke that had nothing to do with any single play.

Watch the Falcons' sideline after they punted from New England's 32-yard line with 17 minutes left. They'd just recovered a failed onside kick. They were still up 19. They had Patriots territory. One field goal likely ends it. The offense that had played with swagger all season started walking to the sideline like a group trying not to make mistakes. The defense began lingering longer between snaps, hands on hips, breathing harder with each tempo play.

Atlanta didn't lose because of Dont'a Hightower's strip-sack in the fourth quarter, though that's the play everyone remembers. The Falcons' defense was on the field for 93 actual snaps (99 total when factoring in penalties). Their pass rush, dominant in the first half without blitzing, evaporated. Atlanta sent an extra rusher on just 7.4 percent of Brady's dropbacks in the first half but was able to generate plenty of pressure with their front four. That pressure disappeared late, with the D-linemen exhibiting rubber legs.

The numbers tell the collapse story that the highlight reel obscures: Atlanta went 1-for-8 on third downs after going 64 percent in the playoffs heading into the game. Their alignment discipline fractured. Their tempo slowed. Their communication broke down—visible in late rotations, linebackers walking up and bailing, defenders peeking inside pre-snap instead of reading their keys.

These weren't random failures. They were signatures of a system under catastrophic stress. And here's the thing: some of this was measurable before each play happened.

The NFL's player tracking system captures location, speed, distance traveled, and acceleration for every player on every play at a rate of 10 times per second, charting individual movements within inches. That's nearly 300 million data points per season, all stored and processed on AWS infrastructure.

NFL Next Gen Stats has used this tracking data to build increasingly sophisticated models. The Coverage Responsibility models now identify which defender was targeted on a given pass and their actual assignment, using frame-by-frame probability calculations that shift as defenders read routes. Tackle Probability models predict the likelihood that a given defender will make a tackle at any given moment, processing over 15 million individual predictions across a season in about an hour using SageMaker Batch Transform.

But tracking dots and biomechanical probabilities still don't capture *stability*. They don't measure the accumulation of micro-signals that predict when a unit is about to fracture. That's where computer vision becomes necessary.

If momentum exists, and the reversal of the hot hand research suggests it does, it's not a binary flip. It's a change in the probability landscape that persists across multiple snaps. Not a one-play spike in win probability, but a sustained shift in how the next sequence behaves: conversion rates, explosive play likelihood, negative play likelihood, penalty likelihood, protection breakdown likelihood, play-calling aggression, time-to-snap, error rates.

This definition makes momentum a *hidden variable* you infer from observable behavior, the way we infer fatigue or tilt in other competitive domains.

The question becomes testable: Can we detect the onset of that persistent shift earlier than humans can?

If big plays aren't reliably momentum triggers, then the play itself isn't the mechanism. The mechanism is whether the system absorbs the shock or fractures from it.

A defense gives up a 35-yard completion. Sometimes they reset. Sometimes they spiral. The difference isn't the yardage. The difference is the system's stability at that moment—substitution readiness, alignment discipline, communication clarity, baseline spacing integrity, quarterback comfort, tempo execution. And emotional containment, which is not mystical. Pose estimation and computer vision can flag subtle changes in form, alerting to mechanical flaws or fatigue before performance drops.

So a momentum engine shouldn't ask "Was that play big?" It should ask "Was the environment brittle?"

Football already has rich numeric data. The NFL tracking system provides location, direction of movement, and speed for every player, essential for predicting trajectories. But momentum isn't only strategy. It's execution quality under stress, and that requires measuring continuous physical reality rather than discrete outcomes.

Machine learning pose estimation models provide accessible, cost-effective, and non-invasive approaches to detailed motion analysis, enabling objective assessment of human movement in clinical settings, sports science, and ergonomics. Computer vision methods can continuously monitor fatigue during athletic performance by predicting both external parameters like generated power and internal parameters like perceived exertion.

Here's what that means for football:

Pre-snap Readiness: Before the ball is snapped, players broadcast intent and condition through stance, weight distribution, and micro-adjustments. A quarterback's base stability—feet width, heel lift frequency, head movement cadence. Offensive line stance consistency, rocking, hand repositioning that correlates with protection call uncertainty. Defensive front tension: forward lean, first-step twitch probability, linebacker depth creep. If you've ever watched a defense look "antsy" before sending heat, that feeling has a measurable signature.

Alignment and Spacing Degradation: Factors like fatigue in combination with pacing and race strategy become possible to study through vision-based approaches, which are not accessible in lab settings. Vision can quantify how far each defender is from their assignment landmark, the variance of defensive spacing across snaps (high variance suggests communication instability or fatigue), the speed and crispness of motion adjustments when the offense shifts. When Atlanta's spacing integrity collapsed in Super Bowl LI, that degradation was visible before the explosive plays happened.

Tempo Physics and Settling Time: Tempo isn't just snaps per minute. It's how quickly a team re-forms into coherence. Time from whistle to huddle break. Jog speed to the line. Time-to-set for linemen, receivers, backs. How often players look to the sideline after getting set; a subtle sign of uncertainty or late play calls. Momentum environments often show asymmetry: one unit is crisp, the other is late. That lateness creates windows.

Micro-Fatigue Signatures: Vision models can analyze gait, stride, and joint angles over time, with slight changes in movement indicating fatigue or imbalance linked to elevated injury risk. Stride length changes. Deceleration profiles between snaps. Time spent with hands on hips or knees. Slow-to-rise duration after contact. First-step acceleration degradation for edge rushers. If you see a pass rusher stop winning the edge by inches, you're watching fatigue. A model can detect it sooner and more consistently than the human eye.

Communication Breakdown: You can't hear calls reliably on broadcast, but you can see confusion. Head-turn frequency and direction clustering—too many defenders looking inward and sideways. Hand signals spiking. Late rotations by safeties. Linebackers walking up and bailing. Offensive linemen turning heads to the center, then back, then down again. These aren't diagnostic in isolation, but when they spike after a previous negative play, they signal organizational stress.

Collision Accumulation: Markerless motion capture systems can measure biomechanical stress with millimeter accuracy, flagging mechanical flaws or fatigue before performance drops or injuries occur. Not all tackles are equal. Some hits change nervous system state even on a two-yard gain. Contact intensity proxies from relative velocity at impact. Whether a ball carrier's torso snaps. Whether tacklers drive through or just tag. Gang tackle accumulation. Whether an offensive lineman is repeatedly knocked backward even if the play is "successful." Momentum can be a physiological tax that shows up several snaps later.

Go back to Atlanta's collapse. The flip didn't happen at the Hightower strip-sack or the Edelman catch. It happened much earlier, when the Falcons punted midway through the third quarter despite leading by 25. The body language shifted. The defense lingered. New England looked fresher, the longer the game went.

A computer vision system monitoring that game would have flagged accumulating instability:

Snap Count Asymmetry: Atlanta's defense approaching the high 80s in plays while New England's tempo forced minimal rest.

Settling Time Degradation: Falcons' defensive alignment speed declining snap by snap.

First-Step Explosiveness Decay: Pass rush win rate dropping not from scheme but from biomechanical fatigue.

Communication Spike: Defensive backs looking inside more frequently pre-snap, late safety rotations increasing.

Spacing Variance Increase: Defenders drifting from assignment landmarks, creating exploitable bubbles.

The Hightower strip-sack didn't create the vulnerability. It exploited a protection scheme already compromised by Freeman's late recognition of the blitz, itself a product of accumulated mental and physical fatigue.

If you label "momentum shift" as "a big play happened," you'll build a big play predictor. That's not useful. You want an engine that says: *the environment is primed for a sequence-level swing*.

A practical labeling scheme: Define a "momentum swing window" as a rolling sequence of 6-10 plays where the probability landscape shifts and stays shifted—sustained win probability delta, sustained expected points added trend change, sustained success rate divergence, sustained increase in negative plays for one team.

Then label the onset as the earliest play where the shift becomes statistically detectable, not the highlight play. This is critical. You're training the model to recognize the ramp, not the explosion.

Now the model's job given the last K seconds of video and the last M snaps of tracking context, predict the probability that a sustained swing begins in the next 1-3 plays.

That's a real predictive target.

A single neural net eating raw broadcast video won't work. Football has relational structure. You need a system that mixes:

Vision Stream: Pose estimation combines ML and computer vision to convert video footage into precise biomechanical data through joint detection and skeleton mapping, with deep learning models like Spatiotemporal Transformers analyzing both spatial and temporal data simultaneously. Extract player keypoints, bounding boxes, field coordinates. Convert into per-player time-series embeddings.

Graph Model: Represent players as nodes with edges defined by proximity, assignment likelihood, formation role. Football is relational. Graphs fit. Using Graph Convolutional Networks, the body is represented where joints are nodes and bones are edges.

Temporal Model: A transformer or temporal convolution network learning patterns across snaps. Momentum isn't a frame. It's a sequence.

Context Encoder: Tabular features for down, distance, score, clock, personnel, formation family, recent play outcomes.

Fusion: Combine into a final predictor with calibration so probabilities mean something. When you output 0.72 momentum risk, it should be roughly true 72% of the time.

You're not building a highlight detector. You're building a stability detector.

Even in deep learning, engineered features help because they encode football priors. An alignment variance score captures how much the defense's pre-snap positions deviate from baseline in similar situations. Settling time measures seconds from break to final alignment. A motion stress score quantifies how often motion causes late shifts or misalignments. Protection uncertainty appears in pre-snap head turns and late pointing among linemen. Pocket deformation tracks how quickly the pocket collapses or drifts.

An explosive posture index aggregates forward lean and first-step twitch probability. Fatigue accumulation combines contact intensity with slow-to-rise duration and between-snap locomotion slowdown. Sideline disorder manifests in substitution turbulence and urgent coach-player interactions. Penalty risk shows up in posture and hand-placement patterns correlated with holds, defensive pass interference, false starts.

None alone is "momentum." Together, they can be signatures of instability.

AWS engineers leverage scalable computing with SageMaker to handle increased data volume, but raw processing power alone isn't enough—it requires deep understanding of football mechanics and extensive feature engineering.

The Big Data Bowl has directly influenced Next Gen Stats development, with previous submissions becoming official stats like Coverage Responsibility this season. If you assume a modern NFL tracking pipeline with frameworks like OpenPose, MediaPipe, and AlphaPose providing high keypoint detection rates, many momentum signatures are computable today: pose-based pre-snap tension via skeleton keypoint analysis, defensive rotation timing and alignment variance across snaps, tempo metrics including time-to-set after motion, micro-fatigue proxies using locomotion classification between snaps, sideline substitution disorder and cluster dynamics, contact intensity proxies using multi-view video and tracking velocities, protection integrity metrics and pocket geometry, penalty risk indicators from posture and hand placement zones.

The hard part isn't "can you measure it." The hard part is "can you define the right label and train a model that generalizes across teams, stadiums, camera angles, and officiating styles while avoiding causality temptation."

Camera Inconsistency: Broadcast angles change, zoom levels change, replays cut away, players occlude each other, sideline shots are intermittent. Methods must derive 3D geometry using constraints like parallel lines with known distances and temporal consistency. Robust tracking across views requires field calibration.

Label Noise: Momentum onset isn't crisp ground truth. Your labeling strategy needs to be defensible and consistent across games, officials, tempo styles.

Domain Shift: Different teams, substitution habits, coaching signals. The model must learn physics and organization, not team-specific quirks.

Causality Temptation: The model will find correlations that aren't causal. That's fine for prediction if it generalizes, but dangerous if you present correlations as explanations. In professional golf research, while a cold hand effect exists, minimal evidence supports a hot hand effect when proper controls are applied, suggesting streaks may be random variation rather than genuine momentum. You need careful interpretation.

If you do this right, you may not even call it momentum in the final product.

You might end up with a Stability Index measuring how coherent a unit is across tempo, alignment, communication, and execution quality under load. Momentum, in that framework, is what happens when stability collapses for one team while it remains high for the other. That would explain why some big plays do nothing. A stable system absorbs shocks. An unstable system fractures from normal contact. And it would explain why some "small" plays flip games. Sometimes the fracture line is already there, and the next ordinary impact is enough.

For thirty years, the hot hand fallacy became an iconic example in behavioral economics of how people are riddled with cognitive biases. It was used to dismiss momentum in domains from sports to financial markets to video gaming. It conflicted with the popular notion of momentum—the idea that athletes' performance can improve after a string of positive results, which psychologists Taylor and Demick in 1994 described as "one of the most commonly referred to and least understood phenomena in the realm of sports."

The reversal of this consensus matters beyond basketball. It suggests we should be more skeptical of academic dismissals of practitioner intuition, especially when practitioners have skin in the game. The beliefs of sports practitioners were in line with reality: the hot hand exists. And this phenomenon has very good reasons to exist from players' strategic behavior.

More importantly for our purposes: there is psychophysiological reality behind momentum. Winners experience psychophysiological responses to success and project their recent victory onto their surroundings by demonstrating aggressiveness, dominance, and determination. Both performers and observers are evolutionarily wired to react to success in competitive situations.

This isn't mysticism. It's biology meeting measurement.

Most fans experience momentum as a moment. A vision model would experience it as a trend. If you build a computer vision engine that treats momentum as a latent stability state, you're no longer hunting highlight plays. You're modeling the environmental conditions that make those highlights inevitable—or impossibly difficult.

That's the deeper claim: football's "flow" is not a story. It's an emergent property of bodies, spacing, tempo, fatigue, communication, and constraint operating under extreme competitive pressure.

Which means it is measurable. And if it is measurable, it is predictable earlier than we currently admit, perhaps early enough to see February 5, 2017 differently. Not as the night Tom Brady pulled off a miracle, but as the night computer vision could have told us Atlanta's defensive stability index was in freefall by the time the third quarter ended.

The question is whether we're ready to measure what practitioners have always known but science spent thirty years denying.

Ehrlich, J. A., Paul, R., Lee, J., & Cain, C. (2025). Momentum in professional golf: A fixed effects reassessment of the hot and cold hand fallacy. Journal of Sports Analytics, 11. https://doi.org/10.1177/22150218251387601 (Original work published 2025)

Miller, Joshua Benjamin and Adam Sanjurjo. “Surprised by the Gambler's and Hot Hand Fallacies? A Truth in the Law of Small Numbers.” Behavioral & Experimental Finance eJournal (2016): n. pag.

“NFL Next Gen Stats.” NFL Football Operations, National Football League, 2026, https://operations.nfl.com/gameday/technology/nfl-next-gen-stats/. Accessed 20 Jan. 2026.

Taylor, J. & Demick, A. (1994). A multidimensional model of momentum in sports. Journal of Applied Sport Psychology, 6, 51-70.