Short answer: yes, AI tennis apps work on clay — but you'll get a noisier report than on a clean hard court, and there are three specific things to do at filming time to get the cleanest possible analysis. Sweep the court (or pick a court that's been swept). Mount the camera on a paved area behind the baseline, not on the clay itself. And keep the camera 6-10 ft above the surface so the freshly-marked white lines are visible from above the dust line.

This post is the longer version of that answer, plus a frank note on where the AI gets confused on clay and what we (and our competitors) actually do about it.

TL;DR

Clay is the hardest surface for AI tennis analysis. Orange ball on orange clay = lower contrast. Dusty/scuffed lines = harder court detection. Both are real, both are fixable in part.
Sweep the lines before you record. White line tape with fresh chalk = court keypoint detection works. Dusty traces = it sometimes doesn't.
Camera 6-10 ft above the surface, behind the baseline. Same as hard courts.
Ballasted tripod on concrete behind the court, not on clay itself. Tripod legs sink into clay over a match — not catastrophic, but visible drift.
Expect ~3-5 F1 points lower shot detection on clay vs hard. This is the published number for AceSense; competitors don't publish theirs.

The honest version: why clay is hard

There's a SwingVision App Store review from a clay-court player that says, plainly: "on clay it doesn't understand where the lines of the court are." That's a real complaint and it's a real problem — not just for SwingVision, but for any AI tennis tool that hasn't specifically trained for clay.

There are two technical reasons clay is harder, and both matter for what you do at filming time.

Reason 1: orange ball, orange clay

A modern tennis ball is fluorescent yellow-green. On a US Open blue acrylic court, it's the highest-contrast object in the frame — it's the only yellow thing against a sea of blue. Frame-to-frame ball tracking (TrackNet on the AceSense pipeline) is essentially solved on this surface.

On red clay, the ball is the same hue as the court, just a slightly different brightness. Add a thin layer of clay dust that the ball gathers after the first bounce, and by the second set the ball is genuinely orange-on-orange. TrackNet still finds it most of the time, but it loses the ball briefly more often — right at the bounce, where you most need it. That bounce moment is what feeds the bounce classifier, which feeds the heatmap.

There's no software fix that fully eliminates this — it's a contrast problem, and contrast is set at the lens. What we can do is augment training data with synthetic low-contrast clay frames so the model doesn't catastrophically fail; that's what we've done. But a hard-court report will always be a hair cleaner than a clay-court report from the same player, on the same camera, with the same software.

Reason 2: the lines are not as crisp as you think

Hard courts have painted lines under acrylic — they don't move, they don't wear, they look the same after 1,000 hours of play. Clay courts have white tape under the surface, with the line edges defined by powdered chalk that gets swept on before each session.

In practice this means three things:

Lines wear during a match. By set three, the baseline is dusted over by ball marks and footwork.
Lines are sometimes doubled. A ball lands on the line, leaves a mark, the chalk drifts — now the court detector sees two parallel lines where there should be one.
Lines look different by light. Bright sun makes the chalk pop; overcast late afternoon, the line edges blur into the clay.

A court-detection model trained mostly on hard courts will sometimes misplace the baseline by 20-30 cm on clay. That's the source of the SwingVision review complaint above.

We've addressed this in two ways: (a) a chunk of our training data is European red and green clay, and (b) we use temporal smoothing — the court keypoints are stabilised across the whole video, not redetected per frame. So even if a frame's lines are dusty, the court geometry from the surrounding 10 seconds carries through.

What to do at filming time

Three concrete things turn a noisy clay report into a clean one.

1. Sweep before you film

If you're filming a match you're playing, you've probably already swept the court — that's standard club etiquette on clay. But specifically: make sure the baseline and singles sidelines are crisply marked. The court-detection model is most sensitive to those four lines. A scuffed centre service line costs you nothing; a scuffed baseline costs you accuracy on bounce localisation.

If you're filming a tournament match you're not playing, you usually can't sweep — but you can pick a court. The freshly-resurfaced courts at the start of the day (or right after a sweep break) give a measurably cleaner report than the same court at 6pm.

2. Mount the camera off the clay surface

This is the bit specific to clay that doesn't apply to hard courts. Don't put the tripod on the clay itself. Two reasons:

Tripod legs sink. Even a 1.5 kg ballasted tripod sinks 1-2 cm into soft clay over a 90-minute match. The camera drifts down. The court drifts up in the frame. By set three, the upper portion of the back baseline can leave the frame entirely.
Court etiquette. Most European clay clubs frown on tripods on the playing surface — you'll get told off.

The fix: most European clay clubs have a paved area behind the court — the path between the back fence and the baseline, or a gravel strip. Set the tripod up there, 6-10 ft behind the baseline. If your club doesn't have one, a length of plywood or even a thick floor mat under the tripod legs works as a portable hard surface.

Full mounting setup is in the no-fence post — most clay clubs do have a fence behind the court, but the fence is often wood or plastic rather than chain-link, in which case a fence clip won't work. Use a tripod.

3. Camera height: 6-10 ft, looking down

Same rule as hard courts, but on clay it matters more — because you want the camera above the dust line. When players are sliding and ball marks are spraying, there's a 1-2 ft layer of fine clay particles in the air around the baseline. A camera at 4 ft of height shoots through that layer. A camera at 7 ft shoots over it.

A 7 ft tripod with the camera mounted on the top, weighted at the base, is the standard clay setup. Cheap (€30-€45) and works for most European clay clubs.

The accuracy hit: how much worse is clay, really?

We measure this. On AceSense's internal benchmark — 200+ test matches across hard, clay, and indoor — the shot-detection F1 score on clay is roughly 3-5 points lower than the same player and camera on a hard court. Most of the gap comes from missed bounces in the half-second after a heavy clay puff (where the ball briefly disappears into the dust cloud).

Stroke-quality scoring is essentially unaffected by surface — the pose model doesn't care what's under the player's feet. Court heatmap accuracy is mildly affected because of the bounce-localisation issue, but the broad pattern (where the player's shots are clustering) is reliable.

Net effect: a clay-court report is useful. It's just a hair noisier than the hard-court version. We publish the full per-surface numbers on the accuracy page; if you want to dig into the methodology, how AceSense works walks through the pipeline.

What about other AI tennis tools on clay?

Honest answer: most are mostly trained on hard courts and report worse on clay than they admit publicly.

SwingVision — the App Store review quoted above is the strongest public signal. Other reviewers report it works "fine" on clay but with worse line-call accuracy. There's no published clay-specific benchmark. If you're an Android player trying to use SwingVision, you can't anyway — see our SwingVision comparison.
PB Vision — paddle-sport-first, tennis support is recent. No known clay benchmark.
Baseline Vision (the €1,800 hardware camera) — designed for hard courts; clay support exists but the camera mounting is awkward on clay clubs.
TennisAI.net — works on clay, no published per-surface accuracy.

Nobody — including us, until our accuracy page is fully public — publishes the per-surface numbers in a way that lets you compare. That's the gap. We're going to fill it.

A note on indoor clay (and other rare cases)

Indoor clay is rarer than indoor hard, but it exists — Italian clubs, German training centres, a few private facilities in France. For those: the camera and mounting rules are the same as outdoor clay, but the lighting issues from the indoor post also apply. Specifically: indoor lighting causes 50-60 Hz flicker on most ceiling lights, which 60fps cameras handle inconsistently. Lock your camera at 30fps for indoor clay if you're seeing flicker bands in the preview.

Bottom line

If you play on clay regularly, you can absolutely use AceSense (or any decent AI tennis app) for self-coaching. You'll get a slightly noisier report than your hard-court friends, but the patterns the report surfaces — shot mix, heatmap zones, stroke-quality scores — are all reliable enough to drive a real practice plan.

The biggest accuracy lever isn't the software. It's whether the camera is 6-10 ft above a swept court, on a stable mount that isn't sinking into the clay. Get those right and the clay-court analysis is genuinely useful.

If you're a club player just getting started, the 5-minute starter guide walks through the first upload end-to-end. And if you're already a SwingVision user fed up with how it handles your home clay club, the side-by-side comparison covers what's different.