Building a smarter eye to read a ball's spin

Spin is one of the most decisive variables in sport, and one of the hardest to measure. It bends a free kick, breaks a curveball, makes a table-tennis serve dip viciously off the bounce. Yet for all its importance, nobody has been able to read it cleanly on a real ball, in a real match — until you stop trying to build a better algorithm and start building a better eye.

Why spin is so hard to see

The magnitudes alone tell you why this is difficult. A table-tennis ball can spin at up to nine thousand revolutions per minute while travelling at thirty-five metres per second. A tennis serve can carry five thousand RPM at seventy metres per second. A golf wedge shot can exceed ten thousand RPM. Spin is central to performance across nearly every ball sport — and brutally hard to capture live.

Conventional frame-based cameras, the kind that grab a full picture sixty or a few hundred times a second, fail here in two ways. First, a ball in flight is tiny in a wide view, spanning only a handful of pixels, so the surface markings you'd need to track rotation are invisible. Zoom in to enlarge it and the field of view narrows so far that you lose the ball almost instantly.

Second, and more fundamentally, the ball moves and spins too fast. High speed across the image causes motion blur even at short exposures, smearing the surface. And the spin rate often exceeds the Nyquist limit of the frame rate — the camera simply isn't sampling fast enough to capture the rotation, so the spin gets aliased and underestimated, the same way a wagon wheel appears to spin backwards in old films. The frame camera is the wrong tool.

A sensor that only reports change

The first ingredient is a different kind of sensor: an event camera. Instead of capturing full frames at a fixed rate, each pixel works independently and asynchronously — it fires a tiny signal the instant the brightness it sees changes, with microsecond timing, and stays silent otherwise. It only reports what changed.

That gives it two superpowers for this problem. Its timing is fine enough that it doesn't alias fast rotation, and it suffers almost no motion blur, because it isn't integrating light over an exposure window. It is tailor-made for something spinning thousands of times a minute.

Lock the gaze, and rotation is what remains

An event camera alone isn't enough, and here is the first beautiful insight. If the ball is hurtling across the sky, the event camera mostly sees that translation — the bulk movement — which drowns out the subtle rotation you actually care about.

So the system adds the gaze-control part of the title. The camera doesn't physically swivel; instead, two fast tilting mirrors — galvanometers — steer the optical line of sight, redirecting where the sensor is looking. These mirrors are quick: they can reposition in about four hundred microseconds and track at speeds equivalent to a ball moving eighty-odd metres per second at a few metres' distance. They keep the ball pinned dead-centre and magnified through a telephoto lens, with a tunable liquid lens holding focus.

It's exactly like following a fast-moving object with your eyes so it stays still on your retina while the background blurs past.

By continuously cancelling the ball's travel across the sky, the mirrors ensure that what's left in the event stream is almost purely the ball's spin.

Turning a smear into a measurement

Now you have a clean stream of events encoding rotation. How do you turn that into a number? The third ingredient is contrast maximisation, and the intuition is lovely. You have a smear of events. You guess a spin — a rotation rate and axis — and you "un-rotate" all the events by that guess, mapping them back to where they'd have come from on the ball's surface. Guess wrong, and the result is a blur. Guess right, and the surface markings snap into a crisp, sharp, high-contrast image. So you simply search for the spin that produces the sharpest image.

The authors add a refinement: they do this on a sphere rather than a flat plane — spherical contrast maximisation — because the ball is round, and projecting onto a flat image causes events to pile up wrongly near the ball's edge and bias the estimate. Doing it on the sphere puts every event at its true location on the curved surface.

That method is accurate but a touch slow for live use, so a final layer makes it fast: a compact convolutional neural network trained to predict the spin directly. Crucially, it is uncertainty-aware — it outputs not just a spin estimate but how confident it is, learned through a loss that lets it express doubt. That confidence is then used to filter and fuse estimates, trusting the certain ones and down-weighting the shaky ones, while a fast parallel search on the GPU refines the magnitude. The network is trained on labels generated by the accurate offline method — so the slow, precise teacher trains the fast student.

The numbers

The results are impressive. Offline, on spinning balls across four sports — table tennis, baseball, tennis and golf — the system measures spin to about two per cent error in magnitude and four degrees in the spin axis. The spherical method beats the previous best contrast-maximisation approach, hitting under two per cent and under two degrees on the test rig, at roughly one-tenth the runtime.

And the headline: live, during actual professional table-tennis play, on the regular ball, it estimates spin with about nine per cent magnitude error and six degrees of axis error — at three milliseconds of latency and seven hundred and fifty measurements per second. Real-time spin, on an unmodified, officially approved ball. It doesn't need a special dotted or marked ball, unlike much prior work; it reads the spin off the ball players actually use.

The honest caveats

The authors are careful. The network's magnitude estimates tend to be conservative at very high spin, simply because there is less training data up there — which is exactly why they bolt on the GPU search to correct magnitude. The biggest practical failure mode is the ball's marking momentarily disappearing: if the logo or texture rotates out of view, there's nothing to lock onto, mitigated by the uncertainty weighting and potentially by using multiple units viewing from different angles.

There is also an honest, almost philosophical caveat about ground truth. In a real match you can't attach a sensor to the ball without changing its flight, so the live accuracy figures are measured against the system's own offline method, not an independent reference — because, for spinning balls in real play, this work is essentially the most accurate measurement available. The live demonstration is table tennis specifically; the other sports were validated on a test rig. And it is specialised hardware — steerable mirrors, a liquid lens, an event camera — though the authors note it is largely off-the-shelf and they plan to release the designs. This is a preprint.

Why it matters

This opens a layer of sports data that has been largely out of reach. The authors point to broadcast analytics, coaching, player development and high-speed object inspection: a live overlay showing the exact spin on a serve, a tool that quantifies a young pitcher's curveball, equipment testing that measures how a racket or club imparts spin. Event tagging and player tracking are mature; the next frontier is high-frequency physical measurement — spin, axis and their uncertainty — captured live, on the real ball.

The deeper lesson is that some problems aren't solved by a better algorithm alone. A frame camera fundamentally cannot see nine-thousand-RPM spin without blur and aliasing, and no amount of clever post-processing fixes a sensor that's looking the wrong way. The breakthrough here is a chain of matched ideas — a sensor that only reports change, an active gaze that cancels motion so rotation stands alone, a sharpening trick that turns a smear into a measurement, and a confidence-aware model fast enough for live play. Each piece enables the next. Sometimes the frontier isn't a smarter model. It's a smarter eye.