How Hawk-Eye works in tennis: a prediction, not a photo

A tennis ball struck hard is in contact with the court for a few thousandths of a second. Nothing in the stadium photographs that contact. The graphic you are shown, the tidy white ellipse sitting half on the line, was not taken from a camera at all. It was computed, from where the ball was in the air before it arrived and from a model of what balls do when they land.

That single fact is missing from most explanations of how the system works, and it is the one that should change how you read every close call you have ever seen overturned. The machine is a predictor. It watches a flight, fits a curve to it, solves for the point where the curve met the ground, then draws a picture of a bounce that nobody recorded.

None of which makes it wrong. It is almost certainly better than the eye of a person standing eight metres away trying to judge a ball moving faster than motorway traffic. But it does mean the confident white shape on the screen is a conclusion, not evidence, and knowing the difference is what separates trusting the technology appropriately from trusting it blindly.

What the camera array actually sees

Start with the hardware, because everything downstream depends on it.

A court fitted for ball tracking has a ring of cameras mounted high around it, in the roof structure, on the lighting gantries, in the stands above the back of the court. They are not broadcast cameras. They are fixed, they do not pan or zoom during play, and their positions relative to the court are measured precisely before the tournament and held constant.

Each camera does one narrow job well: in every frame it captures, it identifies the small bright blob that is the ball and records where in its own flat image that blob sits. That is a two-dimensional answer. A single camera can tell you the ball is thirty per cent of the way across its image and forty per cent of the way up it, which corresponds to a line stretching away from the lens into space. The ball is somewhere along that line. The camera has no idea where.

Two cameras looking at the same ball from different places give two lines, and two lines that both pass through the ball intersect at it. That intersection is the ball's position in the real space of the court. This is triangulation, the same principle a surveyor uses, and it is why the number of cameras matters and why their placement matters more.

More cameras buy two things. They buy redundancy, so a ball hidden from one view by a player's body or a net post is still seen by enough others to be located. And they buy accuracy, because each camera's answer carries its own small error and combining many measurements shrinks the error in the combined answer. A system solving for a position from eight or ten simultaneous views is doing a statistical fit, not a simple crossing of two lines, and the fit tells you not only where the ball probably is but how confident that answer is.

The cameras run fast, far faster than broadcast frame rates, because a serve crosses the court in well under half a second and a system sampling at television speed would get only a handful of positions during the whole flight. High frame rates are the difference between fitting a curve to forty points and fitting it to six.

67%33%

• Ball tracking for line calls12
• Foot fault monitoring6

The configuration used for the automated line-calling installation at the 2020 US Open. Twelve cameras track the ball for line calls; a further six feed the official who monitors for foot faults.

Show the numbers

The 18 cameras of a fully automated court, by job

Item	Value
Ball tracking for line calls	12
Foot fault monitoring	6

The foot fault cameras deserve a sentence of their own, because they are the part of the installation nobody talks about and the part that most changed the experience of serving. A line judge standing at the baseline could see a foot fault only from one angle and only if looking at the right moment. A dedicated camera set watching the baseline sees every serve from several angles and never gets bored. Players who had spent careers just about getting away with a creeping front foot found out about it very quickly.

The bounce is a prediction, not a photograph

Here is the mechanism that the marketing graphics obscure.

The cameras see the ball in the air. They do not see it on the ground, or rather they see so little of the contact that the contact itself cannot be measured directly. At a few thousand frames per second you might catch one frame with the ball touching down, and one frame is not a measurement of anything.

So the system does not try to observe the bounce. It reconstructs it.

Take the last stretch of the ball's flight before it lands. The system has a sequence of three-dimensional positions, each with a timestamp. Those positions describe a curve through space, and that curve is not arbitrary: it is the path of a spinning sphere under gravity and air resistance, which is a shape physics can describe with a small number of parameters. Fit the parameters to the observed points and you have a continuous flight path, defined everywhere, including in the gaps between frames where no camera saw anything.

Now solve for the moment that path reaches the height of the court surface. That gives an instant in time and a point on the ground. Not a photographed point. A calculated one.

The distinction is not pedantry. It has three practical consequences. The first is that the quality of the answer depends on the quality of the fit, which depends on how many clean positions the cameras got during the approach, which depends on whether the ball was occluded, badly lit, or moving so fast it smeared across the sensor. A short, well-seen approach gives a superb answer. A ball hidden behind a player until the last instant gives a worse one. The screen graphic looks identical either way.

The second consequence is that the system is extrapolating a small distance beyond its last observation, and extrapolation always costs more accuracy than interpolation. The last camera position before the bounce might be a few centimetres up in the air. The answer at the ground is a projection from there.

The third is the one everyone should know and almost nobody does. The picture you are shown is drawn after the decision is made, for the benefit of the audience. It is a rendering of the system's conclusion, not the input to it. When you look at that ellipse and feel that you can see the ball touching the line, you are looking at an animation of a number.

Why the mark is an ellipse and not a dot

A ball does not touch the court at a point, and the shape of the mark it leaves is one of the most useful things to understand about line calling.

Two things happen during contact. The ball flattens, because a pressurised rubber sphere hitting a hard surface at speed compresses against it, and the flattened region is a disc rather than a point. And the ball keeps moving forward while it is flattened, because it arrived at a shallow angle with most of its speed horizontal, so the compressed patch slides along the surface before the ball recovers its shape and leaves.

Compression alone would give you a circle. Compression plus slide gives you a smear stretched along the direction of travel. That is the ellipse.

The proportions change with the shot. A fast flat serve arrives at a very shallow angle carrying enormous horizontal speed, so it slides a long way while flattened and leaves a footprint several times longer than it is wide. A high, heavily topspun ball drops in steeply, digs in, and grips rather than slides, so its footprint is shorter and rounder. Backspin does something else again, holding the ball on the surface and lengthening the contact.

This matters for calls because a ball is in if any part of it touches any part of the line. The relevant question is therefore not where the centre of the ball was, it is whether the leading edge of the contact area reached the outer edge of the line. On a serve landing long by a whisker, the difference between in and out is the difference between the front of a long smear and the back of it.

It also explains the calls that look wrong on television. A ball shown as in, with only the tip of the ellipse overlapping the line and the bulk of the shape well beyond it, looks to the eye like a ball that was out. It was not. The system is telling you that contact began on the line and continued past it, which is precisely what the rules count as in. The graphic is honest. Our intuition about it is not.

Understanding this also tells you why the technology has always been more comfortable on hard courts and grass than on the surface where a bounce leaves physical evidence, which is a fight we will come back to.

How Hawk-Eye works in tennis, in the order it happens

Everything above happens between the strike and the call, in less time than it takes a player to turn round.

1. 1The ball is struckEvery camera in the ring is already running and already synchronised to a common clock, so each frame from each camera carries the same time reference. Nothing needs to be started.
2. 2Each camera finds the ball in its own frameSoftware isolates the ball from the background in every image and records its position within that image. This is a flat, two-dimensional answer, and on its own it locates nothing.
3. 3The views are combined into a positionBecause the exact position and orientation of every camera relative to the court is known, the separate views are solved together into one point in three-dimensional space. Repeating this for every frame produces a string of positions with timestamps.
4. 4A flight path is fittedThe positions are fitted to the physical curve a spinning ball follows under gravity and air resistance. The result is a continuous path defined even between the frames, with the noise in individual measurements smoothed out.
5. 5The impact is solved forThe path is followed forward to the height of the court surface. This gives the instant of contact and the point of first contact, neither of which any camera saw directly.
6. 6The footprint is modelledUsing the ball's speed, angle and spin at impact, the system works out how far the ball compresses and how far it slides while compressed. That produces the contact area, an ellipse stretched along the direction of travel.
7. 7The footprint is tested against the lineThe court has been surveyed, so the exact position of every line is known in the same coordinate system as the ball. If any part of the contact area overlaps any part of the line, the ball is in.
8. 8The call is madeUnder live calling a recorded voice announces out or fault through the court speakers, loud and immediate. The graphic is generated afterwards, for the screen and the broadcast, from the decision already taken.

The order is fixed. Every step is complete before the next begins, and the whole chain runs in a fraction of a second.

Read the last two steps together. The court is surveyed and the ball is calculated, and the call is the comparison of one against the other. If either the survey or the calculation drifts, the call drifts with it, silently, and nothing on screen looks any different. That is the whole argument for calibration, which we will get to.

What a margin of error of a few millimetres actually means

The manufacturer's published figure is a mean error of around 3.6 millimetres. It has been quoted for years and it is the number every discussion of accuracy comes back to.

Three things about it are worth pinning down.

It is a mean, not a maximum. A mean error describes the middle of a distribution of errors, which means individual calls are sometimes better and sometimes worse. A system with a mean error of a few millimetres will occasionally produce a call that is off by more, in the same way that an average of anything is not a promise about every instance of it.

It is small compared with a ball. A tennis ball is roughly 67 millimetres across. An error of a few millimetres is a small fraction of that, and a small fraction of the width of a line. For the overwhelming majority of calls, an error of that size cannot change the answer, because the overwhelming majority of balls land nowhere near a line.

And it is not zero, which is the part that matters. When a ball is shown touching the line by a fraction of a millimetre, the system is reporting a difference smaller than its own uncertainty. It has still made a decision, because the rules require one, and the decision is final. But the honest description of that call is not that the ball was in. It is that the best available estimate places it marginally in, with a tolerance wide enough to have gone the other way.

• 3.6Published mean error, in millimetres
• 3Challenges per set under the standard rule
• 1Extra challenge granted for a tiebreak
• 1Grand Slams still using human line judges in 2026

The error figure is the manufacturer's long-published mean. The challenge allowances are set by the tour rules. The Grand Slam count is the position for 2026.

Why does a margin exist at all? Because every step in the chain adds a little. The cameras have finite resolution, so the ball's position in each image is known to within a pixel or so, and a pixel at that distance is not nothing. The synchronisation between cameras is very good but not perfect, and a ball at serve speed covers real distance in a microsecond of timing slop. The curve fit smooths noise but also imposes a model, and any model is an approximation of a real flight through real air. The impact solve extrapolates. The footprint model uses estimated spin, and spin is the hardest of the flight parameters to measure well.

Each of those contributes a fraction of a millimetre, and the combined figure is what testing measures.

Here is a constructed example to show why the tolerance is not academic. Invented figures, chosen to be round rather than realistic. Suppose a serve arrives with a horizontal speed of 50 metres per second. In one millisecond it travels 50 millimetres. If the timing of the moment of impact were uncertain by a fifth of a millisecond, the position of the impact point would be uncertain by 10 millimetres in the direction of travel. That is more than the published mean error on its own, which tells you the real system must resolve timing considerably better than a fifth of a millisecond. The arithmetic is invented; the point it illustrates is not, which is that at these speeds tiny timing errors become visible distances.

The challenge system was a truce, not a solution

Before the machine called anything, it argued with the humans on request.

The ITF signed off on the technology in 2005 after testing. In March 2006 it was used officially for the first time at a tour event, in Miami, and later that year the US Open became the first Grand Slam to run it. Players could ask for a review of a call, a limited number of times, and the review was binding.

The allowance settled at three unsuccessful challenges per set, with an additional one granted for a tiebreak, and the count reset at the start of each new set. A successful challenge cost nothing: get it right and you keep the challenge. Get it wrong and you lose one. The design was deliberate, and it turned challenging into a small strategic game of its own, with players hoarding their last one for the end of a set and spending them freely early.

1. 1The call is madeA line judge calls the ball out, or says nothing, and the chair umpire either accepts that or overrules it. The point either continues or stops.
2. 2The player asks, at onceThe challenge has to come immediately, before the player prepares to serve or receive the next point. A challenge raised after the game has moved on is refused, which is why players learned to raise a hand before they had finished thinking.
3. 3The umpire confirms what is being challengedThe umpire states the call under review and the score, out loud, so there is no argument afterwards about which ball was being questioned. The clock effectively stops.
4. 4The replay is played to the whole stadiumThe animation is shown on the big screen and to the television audience at the same time, with no private conversation and no delay while a decision is reached elsewhere. The crowd usually claps along.
5. 5The result standsThe system's answer is final and the umpire announces the corrected score. If the challenge was successful the player keeps their allowance; if not, one challenge is gone.
6. 6Play resumesWhere a rally was stopped by a wrong call that the review reverses, the point is replayed rather than awarded, unless the player would clearly have won it outright.

The sequence had to be immediate. A player who played the next point had accepted the call, whatever they said afterwards.

The challenge system did something no rule change had managed before: it turned officiating into entertainment. The suspense of the replay, the crowd clapping in rhythm, the slow reveal of the ellipse, all of it became part of the show. Broadcasters loved it. Players hated losing challenges and loved winning them.

It also had a defect that everyone noticed and nobody could fix. Under a challenge system, most wrong calls stay wrong. A player only gets a correction if they choose to spend a challenge, and they choose based on their own view of a ball travelling at speed away from them, which is exactly the judgement the system exists to replace. Balls called out that were actually in went uncorrected all the time, because the player did not think it was worth a challenge, or had run out, or simply did not see it.

The technology was capable of calling every ball from the day it was installed. The challenge system chose not to let it, because the sport was not yet ready to hand the decision over. That is the honest description of the compromise: it kept humans nominally in charge while quietly proving them wrong a few times a set. The detailed rules governing what could be challenged, when, and with what consequence are worth reading on their own, and the mechanics of the challenge allowance explain several oddities that survived into the live-calling era.

Live calling: the day the line judge stopped mattering

The first tournament to remove the intermediate step and let the system call everything was the Next Gen ATP Finals in 2017, an event that exists partly to test ideas the tour is not ready to impose. It worked. The calls were instant, they were spoken through the court speakers, and the arguing stopped.

From there the spread was quick by the standards of a conservative sport. The US Open used automated calling on most courts in 2020 and made it permanent in 2022. The Australian Open ran a Grand Slam entirely on electronic calling in 2021. The ATP committed to electronic line calling at all its tour events from 2025, and Wimbledon, of all tournaments, dropped its line judges the same year.

That last one is the one that told you the argument was over. Wimbledon dressed its line judges in specially designed uniforms as part of the tournament's visual identity, and getting rid of them was a genuine cultural loss that the club decided was worth taking. The reason given was consistency: the tour was calling every ball electronically, so a player arriving at a Grand Slam should not be asked to switch back.

What was actually lost is more interesting than the uniforms. A line judge is a decision-maker with a voice, and their calls could be overruled by the chair umpire, which meant a close call had two humans on it and a visible process for resolving disagreement between them. Live calling removes both. The call arrives already final, from a speaker, in a recorded voice, and there is nobody to look at.

Players noticed the strangeness of it immediately. The stimulus that used to say "stop, that was out" now comes from a machine that never hesitates, never varies its tone and never gets it slightly wrong in a way you can hear. Several have said the silence between points feels different. That is not a complaint about accuracy. It is a complaint about theatre, and theatre is a real part of a live sport.

The change also removed one of the last places where a match could be genuinely disputed, which has knock-on effects for the way a scoreboard turns points into pressure. A break point saved by a call the receiver believed was wrong used to be an event with an afterlife: the player carried it, the crowd carried it, the commentary carried it for an hour. Now it is a fact, announced and closed, and the match moves on.

Clay had a rival technology already, and it was made of dirt

Roland Garros is the only Grand Slam that still uses human line judges, and it confirmed it would keep them for 2026. That looks like stubbornness from outside. It is not, or not only.

Clay is the one surface that keeps its own record. A ball landing on crushed brick leaves a mark, and that mark is physical evidence available to anyone who walks over and looks at it. The chair umpire's authority on clay includes climbing down and inspecting the mark, and the procedure for doing so is old, well understood, and accepted by players who have grown up on the surface.

So on clay the technology is not filling a vacuum. It is competing with an existing method that has an unusual property: it is a direct observation of the actual bounce rather than a reconstruction of it. Every other surface offers nothing at all, and a calculated answer beats no answer easily. On clay a calculated answer has to beat a real mark, and that is a harder sell.

There are complications on the clay side too, which is why the argument is not as one-sided as it sounds. A mark has to be the right mark, and courts accumulate marks by the dozen. The umpire has to identify which one the disputed ball made, which involves judgement, and players dispute the identification as often as they dispute the call itself. The surface moves under players' feet, so a mark can be smeared or partly erased before anyone reaches it. And a ball that lands and skids on loose clay can leave a mark that overstates or understates where the contact actually began.

The French federation's stated objection has been reliability on a surface that changes constantly with weather and play, and the observation that it employs officials whose job this is. Both are real. The tournament also has an institutional preference for the human version of the sport that it makes no secret of.

Where this ends is probably obvious. Every other significant event has moved, the tour requires the technology above a certain level, and a single holdout is not a stable position over a decade. But the clay argument is the most substantive objection anyone has raised, because it is the only one that offers an alternative rather than a preference. The peculiarities of the surface are worth understanding on their own terms, and what clay does to a bouncing ball is the root of most of them.

Calibration is the invisible half of the job

Everything the system knows about the court comes from a survey, and everything it knows about itself comes from calibration. Neither is a one-off.

Before an event, the court is measured precisely and the positions of the lines are recorded in the same coordinate system the ball will be tracked in. The cameras are then calibrated: their exact positions, orientations and lens characteristics are determined, usually by observing known reference points around the court, so that the software can convert a pixel in an image into a line in space with the correct geometry.

Things move. Stadium structures flex with temperature. A camera bumped by a maintenance crew is a camera pointing somewhere slightly different. Lighting changes across a long day session, and a ball looks different against a bright background than a dark one. Systems handle this by re-checking their calibration continuously against fixed features of the court, and by flagging cameras whose view no longer agrees with the others.

Validation is the other half, and it is what makes certification meaningful. Independent testing of a line-calling system involves firing balls at known positions, with known speeds and spins, and comparing the system's answers with the measured truth. That is how a claimed error figure becomes a verified one, and it is why the ITF's sign-off in 2005 mattered more than any manufacturer's claim. The tolerance is not something the operator declares. It is something a test rig establishes.

None of this is visible from a seat in the stadium, which is the problem. A miscalibrated system produces confident wrong calls that look exactly like confident right ones. There is no equivalent of watching a line judge and thinking they are having a bad day.

What Hawk-Eye in tennis cannot do, and what happens when it stops

The system calls lines. It does not call anything else, and the list of things outside its remit is longer than people assume.

It does not judge a double bounce. Whether a ball bounced twice before a player reached it is a call the chair umpire makes with their own eyes, and it is one of the harder calls in the sport, made at ground level with a player's body in the way. It does not judge whether a player touched the net, or whether the ball was struck before it crossed the net, or whether a player was hindered. It does not adjudicate a service let unless a separate net sensor is fitted, which is a different piece of equipment doing a different job. And it decides nothing about conduct, timing or the many judgement calls the chair is there for.

Then there is failure. In 2025 at Wimbledon the ball tracking was switched off on one side of the court during a match, by human error rather than a technical fault, and three calls went unmade in the time it was down. The chair umpire made two of them without knowing the system had been deactivated. When the problem was discovered the point was replayed, which satisfied nobody, and the tournament apologised to both players. The club's response was to remove the operators' ability to deactivate the tracking manually, which is the correct engineering answer: if a human can turn it off by mistake, take away the switch.

The episode is more instructive than an embarrassment usually is, for three reasons.

The fallback is one person. Without line judges on court, a failure hands every line on both sides of the net to the chair umpire, who sits at the middle of one sideline and was not hired to call baselines. The redundancy that a full complement of officials used to provide has been removed on purpose, and the sport has accepted a lower floor in exchange for a higher average.

Nobody knew it was down. The system's failure mode was silence, and silence in tennis means the ball was in. A machine that stops calling looks identical to a machine that has decided every ball is good, which is close to the worst possible way for a safety-critical system to fail. Engineers have a word for the opposite property, fail-safe, and a line-calling system that goes quiet is not it.

And the remedy is thin. Once a call has been missed and play has continued, the options are to replay the point or to let it stand. Neither restores what was lost, and the choice is made by an umpire under pressure with an incomplete picture. Set that against the deliberate, structured review process that football's video officials work through, which has plenty of problems of its own but at least defines a procedure for putting a specific kind of mistake right.

Did removing all the doubt make the sport better?

The accuracy question is settled and not very interesting. A calibrated camera array beats a human eye on a ball travelling at motorway speed, and anyone arguing otherwise is arguing about something else while pretending to argue about accuracy.

The interesting questions are the ones underneath.

Certainty is not the same as fairness. A system with a small tolerance still makes calls inside its tolerance, and it makes them with total confidence and no appeal. Under the old arrangement, a genuinely fifty-fifty ball was resolved by a human who might be wrong, and everyone understood that. Now it is resolved by a calculation that might be wrong, and nobody understands that, because the graphic looks certain. Replacing visible fallibility with invisible fallibility is a real change and it is not obviously an improvement in fairness. It is an improvement in the appearance of fairness.

The sport lost an argument it used to enjoy. Disputed calls generated conflict, and conflict generated characters. A player berating an umpire over a line call is a scene that has produced some of the most watched moments in tennis history. Those moments are gone, and nothing has replaced them, because you cannot shout at a speaker.

Officials lost a career step. Line judging was the entry level of tennis officiating, the job people did on the way to becoming chair umpires, and removing it removes the pathway. This is the least discussed consequence and probably the most serious one over twenty years. A sport that stops training officials at the bottom eventually runs short of them at the top, which is a problem the wider professional tennis structure will have to solve deliberately rather than by accident.

Players got something they wanted. The dominant player view has been positive, and it is easy to see why. A wrong call at a break point can decide a match and a career, and reducing those is worth almost any amount of atmosphere. Players are not paid to protect theatre.

And the game got faster in a way nobody planned. Instant calls with no challenges, no replays and no arguments strip a measurable amount of dead time out of a match. Whether that is a gain depends entirely on whether you thought the dead time was dead.

How to watch a line call now, and what to notice

Four habits will tell you more about what you are seeing than any amount of staring at the ellipse.

Notice when the graphic is not shown. Under live calling, most calls are announced and never illustrated. A replay appears when a broadcaster asks for one, usually because the call was close or a player reacted. If you are seeing a picture, someone has already decided the call was interesting, which is a filter on your view of the system's accuracy.

Watch the shape, not the position. A long, thin footprint tells you the ball arrived fast and flat and slid a long way. A short, round one tells you it dropped in steeply with topspin. The shape is the system showing you its working, and it is the part of the graphic that carries real information about the shot that produced it.

Treat a millimetre call as a coin toss that has already landed. If the overlap is a sliver, the honest reading is that the ball was on the boundary of what the system can resolve. The call stands, because a call has to. It does not mean the ball was in by a millimetre. It means the estimate says so.

Ask what the fallback is. At an event without line judges, the answer is one person in a chair. At Roland Garros, the answer is a set of officials and a mark in the clay. Neither is better in the abstract. They fail differently, and knowing which failure you are exposed to is the useful thing.

The technology did what it was asked to do. It removed the routine wrong call, and it did so completely enough that a generation of players will finish their careers without ever losing a match to one. What it could not do is remove uncertainty, because uncertainty is a property of measuring a fast small object against a painted line, not a property of the people who used to do the measuring. The uncertainty is still there. It just stopped being announced.