On ASPset-510, a public CC0 three-camera dataset, Autelica syncs cameras from the athlete’s motion to a 1.2 ms median error across 35 clips and measures top speed within 2.9% median of the dataset’s 3D reference on 12 running clips when the athlete’s height is known. On our Sep 30 2026 three-drone flight, camera orientations solved from the footage agreed with the drones’ logs to 0.3°. The weakest result is image quality from a hidden drone’s viewpoint: 14.9 to 15.1 dB PSNR against about 21 dB when the hidden camera is used.
01Results
| What | Result | Measured on | Against | Date |
|---|---|---|---|---|
| Robot, own drone footage | 60 / 60 physics runs per captured segment, 59 with random pushes, floor changes and delays; joint error 4.2–5.9° | Three segments from the Sep 30 2026 three-drone flight, retargeted to a Unitree G1 | The captured motion | Oct 2026 |
| Robot, three cameras vs dome | 59 / 59 runs both; keypoint error 12.9 vs 12.3 cm | A tennis clip from the CMU Panoptic dome (evaluation only, non-commercial) | The dome’s dense reconstruction | Oct 2026 |
| Camera sync from motion | 1.2 ms median error, 2.4 ms at the 90th percentile | 35 clips of 4 s or more, ASPset-510 (three fixed 4K cameras) | The dataset’s wired hardware sync | Oct 2026 |
| Sync of two phones | Recovered in 16 of 16 clips | Two iPhones whose file clocks were up to 2.1 s apart | Known offsets | Sep 2026 |
| Camera orientation | 0.3°; positions within 0.2 to 0.4 m | Three DJI Neo 2 drones, Sep 30 2026 flight | The drones’ own flight logs | 30 Sep 2026 |
| Top speed | 2.9% median error with the athlete’s height known; about 8% low with an average adult height assumed | 12 running clips, 2 athletes, ASPset-510 | The dataset’s 3D reference | Oct 2026 |
| Worked example | Top speed 17.8 vs 18.1 km/h; distance 21.33 vs 21.34 m | One 9-second sprint, ASPset-510 (pipeline as of Sep 2026) | The dataset’s 3D reference | Sep 2026 |
| Held-out camera, joints | About 6 to 9 pixels on a 4K frame | A sprint from ASPset-510, joints rebuilt from the other two cameras | Joints detected in the hidden camera | Oct 2026 |
| Held-out drone, image | 14.9 dB PSNR (take E), 15.1 dB (take F), rendered with no real pixels from the hidden drone | 20 s from each of two takes, Sep 30 2026 flight | The hidden drone’s footage; about 21 dB when the renderer may use that drone | Oct 2026 |
| Processing time | About 40-45 s for a 6 s three-camera 4K clip including the replay reel; about 20 s for one camera; 7.4 min for 30 s of three-drone footage | One RTX 5090 desktop GPU | Wall-clock processing time, inputs on disk to finished replay | Oct 2026 |
SourcesASPset-510 · arXiv 2108.03936. Checked 2026-10-02.
02Method
Sync. Clocks are lined up first from sound, then from the athlete’s 2D joints, by finding the time shift that makes the three views agree in 3D. On ASPset-510 the answer is checked against the dataset’s wired sync.
Cameras. Each camera’s lens and position are solved from the images and the athlete’s body. On the drone flight we compare the solved orientations and positions with the drones’ flight logs. The logs place cameras well but their timestamps were off by half a second to a second, so we sync from the picture.
Speed and distance. A parametric body is fitted to all cameras every frame; speed comes from the body’s path. Scale comes from the athlete’s height, which is why an assumed average height reads about 8% low.
Images. One camera is hidden, its view is rendered from the others and compared with the real frame in PSNR. The same held-out check grades every clip we deliver.
03What these numbers don’t show
- ASPset-510 has fixed cameras and a known set-up. Moving drones are harder, and our drone numbers come from one day of flying.
- The speed result covers 12 clips from 2 athletes. It is not a validation study.
- The held-out image scores are low. Views far from any real camera show softness and holes.
- No independent party has reproduced these results yet.
04Data and licences
ASPset-510 was filmed by the Australian Institute of Sport and La Trobe University and released under CC0. The drone footage is our own, flown three drones, one pilot each, under FAA Part 107. We use non-commercial datasets and model weights only for benchmarks.
05Changelog
- 3 Oct 2026, eveningMoved camera calibration and sync refinement onto the GPU. Calibration is more accurate as well as faster: top-speed error with the athlete’s height known is now 2.9% (about 8% low without it), on the same 12 public clips, and a three-camera job takes about 35 seconds on one GPU.
- 3 Oct 2026, laterShipped a new release: a keypoint detector fine-tuned for people in gear, with flip testing, plus faster processing. On the same public clips, sync is now 1.2 ms median (2.4 ms p90), joint misses 6-9 px, and top-speed error about 5% with the athlete’s height known (was 3.6%; about 12% low without it). We publish what the product runs.
- 3 Oct 2026Re-ran every number with commercially licensed detectors and masks only (D-FINE, HRNet-W48, SAM 2.1). Sync improved from 1.3 to 1.2 ms median; joint misses moved from 6-11 to 7-10 px; held-out drone image scores are now quoted with no real pixels from the hidden drone (14.9 and 15.1 dB, previously 15.0 and 15.7 dB with a looser protocol).
- 2 Oct 2026Page published.
Questions
How accurate are the 3D joints?
Body joints are within about 52 mm of the three-camera outdoor sports reference in ASPset-510, and hands within 17–19 mm indoors at about 2 m on Harmony4D (evaluation only).
How is each clip graded?
One camera is left out of the solve and the 3D is compared with what that camera saw, in pixels and in millimetres at keypoint depth. That grade ships with every clip.
Has this run on a real robot?
Not yet. All robot results are in MuJoCo physics with randomised pushes, floor changes and control delays.
How accurate is Autelica’s camera sync?
1.2 ms median error and 2.4 ms at the 90th percentile against wired hardware sync, on 35 clips from the public ASPset-510 dataset.
How accurate is speed measured from three cameras?
On 12 running clips from ASPset-510, top speed was within 2.9% median of the dataset’s 3D reference when the athlete’s height was known, and about 8% low with an average height assumed.
How good does a replay look from an angle no camera filmed?
Measured from a hidden drone, 14.9 to 15.1 dB PSNR, against about 21 dB when the renderer may use that drone. Views close to the real cameras look much better than views far from them.