# Soeiro-Vuong and Rochet test tennis event detectors beyond broadcast footage [Canonical reading](https://escapemeridian.com/readings/soeiro-vuong-and-rochet-test-tennis-event-detectors-beyond-8143d01d) David Soeiro-Vuong and Andre-Louis Rochet argue that tennis event detectors built around broadcast footage can fail when the camera moves elsewhere. Their benchmark covers ball bounces and racket hits across broadcast, fixed elevated, phone and court-level views, where bounce and hit are easy to confuse. Supervised spotters using published broadcast weights scored near-perfectly on broadcast footage and almost nothing on phone and court-level footage. RallyArb, an arbitration layer over a ball track and player poses, kept bounce F1 above 0.8 on three phone and court-level clips, but two were used to tune it; on the held-out clip it found 8 of 11 bounces. It placed hits too far from the contact frame to match human timing. The vision-language models detected events across viewpoints but missed the contact frame by about 0.1 s. The authors also report that fine-tuning the spotter reached about 0.8 bounce F1 on two clips but stayed at 0.2 on a grazing phone angle. The benchmark, evaluator and frozen predictions are released, making the reported comparisons inspectable. The results show a gap between recognizing events across camera views and timing them accurately; how well these methods transfer beyond the tested clips remains an open question. Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims. ## Original source arXiv cs.AI · Article Original source date: Oct 27, 2026 (2026-10-27). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-13T04:18:25.924Z. First observed: 2026-10-30T06:40:46.035Z. Retrieved: 2026-10-31T08:04:28.771Z. - [Can Event Detectors Leave the Broadcast Camera? Evidence from Tennis](https://arxiv.org/abs/2610.38042) ## Limits The benchmark contains eight clips, and RallyArb was tuned using two of the three phone and court-level clips on which its bounce F1 is reported. The abstract does not establish how the reported results generalize beyond the tested clips. ## Open question How would these results change on additional phone and court-level clips that were not used to tune RallyArb? ## Source-stated The authors introduce a benchmark for tennis bounces and hits across several camera viewpoints, with annotations, an evaluator and frozen predictions. > We release a benchmark of 345 bounce and hit annotations at frame level on eight clips that mix broadcast, fixed elevated, phone and court-level viewpoints - [Supporting source](https://arxiv.org/abs/2610.38042) ## Source-stated The authors report that supervised spotters with published broadcast weights perform well on broadcast footage but poorly on phone and court-level footage. > The supervised spotters score near-perfectly on broadcast footage and almost nothing on phone and court-level footage - [Supporting source](https://arxiv.org/abs/2610.38042) ## Source-stated RallyArb achieved bounce F1 above 0.8 on three phone and court-level clips, although two were used to tune it; it found 8 of 11 bounces on the held-out clip. > RallyArb keeps a bounce F1 above 0.8 on the three phone and court-level clips, two of which were used to tune it, and finds 8 of 11 bounces on the held-out one - [Supporting source](https://arxiv.org/abs/2610.38042) ## Source-stated The authors say RallyArb's hit timing was too far from the contact frame to match human timing. > it places hits too far from the contact frame to match human timing. - [Supporting source](https://arxiv.org/abs/2610.38042) ## Source-stated The vision-language models detected events from every viewpoint but missed the contact frame by about 0.1 s. > The vision-language models detect events on every viewpoint and miss the contact frame by about 0.1 s. - [Supporting source](https://arxiv.org/abs/2610.38042) ## Source-stated The authors report that fine-tuning the spotter reached about 0.8 bounce F1 on two phone and court-level clips but remained at 0.2 on a grazing phone angle. > Fine-tuned leave-one-clip-out from its broadcast weights, the spotter reaches about 0.8 bounce F1 on two phone and court-level clips and stays at 0.2 on a grazing phone angle. - [Supporting source](https://arxiv.org/abs/2610.38042) ## Source and contributor context Soeiro-Vuong, David, Rochet, Andre-Louis · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media No suitable media was found by the supported page extraction. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Benchmark A defined set of tests used to compare systems. Results depend on the tasks and measures selected. - [Definition reference](https://arxiv.org/abs/2211.09110) ## Fine-tuning Further training of an existing model to adapt it to a particular task or use. - [Definition reference](https://developers.google.com/machine-learning/glossary#fine-tuning) ## arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. - [Explore this context](/entities/arxiv) - [Reference](https://arxiv.org/) - [Reference](https://info.arxiv.org/help/api/index.html) ## Editorial provenance Version 1. Publication ID: c31c793f-7ac6-4779-811d-57879f5d97b8. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-13T04:18:25.924Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: bc027260bdc6a9f4594fe6615e50f01d1f248eb9ec4d31ae090fd49bdbcf30ef. Source text hash: 0b36ee7926c0984f5787aeb16d607f79a56ee600663c6f9578ffbdb03bd9eaea. ## Revision and source-check history ### Version 1 — 2026-11-13T04:18:25.924Z David Soeiro-Vuong and Andre-Louis Rochet argue that tennis event detectors built around broadcast footage can fail when the camera moves elsewhere. Their benchmark covers ball bounces and racket hits across broadcast, fixed elevated, phone and court-level views, where bounce and hit are easy to confuse. Supervised spotters using published broadcast weights scored near-perfectly on broadcast footage and almost nothing on phone and court-level footage. RallyArb, an arbitration layer over a ball track and player poses, kept bounce F1 above 0.8 on three phone and court-level clips, but two were used to tune it; on the held-out clip it found 8 of 11 bounces. It placed hits too far from the contact frame to match human timing. The vision-language models detected events across viewpoints but missed the contact frame by about 0.1 s. The authors also report that fine-tuning the spotter reached about 0.8 bounce F1 on two clips but stayed at 0.2 on a grazing phone angle. The benchmark, evaluator and frozen predictions are released, making the reported comparisons inspectable. The results show a gap between recognizing events across camera views and timing them accurately; how well these methods transfer beyond the tested clips remains an open question. ## Original source arXiv cs.AI · Article Original source date: Oct 27, 2026 (2026-10-27). Date basis: Date basis not provided. Added to Escape Meridian: 2026-11-13T04:18:25.924Z. First observed: 2026-10-30T06:40:46.035Z. Retrieved: 2026-10-31T08:04:28.771Z. - [Can Event Detectors Leave the Broadcast Camera? Evidence from Tennis](https://arxiv.org/abs/2610.38042) ## Limits The benchmark contains eight clips, and RallyArb was tuned using two of the three phone and court-level clips on which its bounce F1 is reported. The abstract does not establish how the reported results generalize beyond the tested clips. ## Open question How would these results change on additional phone and court-level clips that were not used to tune RallyArb? ## Source-stated The authors introduce a benchmark for tennis bounces and hits across several camera viewpoints, with annotations, an evaluator and frozen predictions. > We release a benchmark of 345 bounce and hit annotations at frame level on eight clips that mix broadcast, fixed elevated, phone and court-level viewpoints - [Supporting source](https://arxiv.org/abs/2610.38042) ## Source-stated The authors report that supervised spotters with published broadcast weights perform well on broadcast footage but poorly on phone and court-level footage. > The supervised spotters score near-perfectly on broadcast footage and almost nothing on phone and court-level footage - [Supporting source](https://arxiv.org/abs/2610.38042) ## Source-stated RallyArb achieved bounce F1 above 0.8 on three phone and court-level clips, although two were used to tune it; it found 8 of 11 bounces on the held-out clip. > RallyArb keeps a bounce F1 above 0.8 on the three phone and court-level clips, two of which were used to tune it, and finds 8 of 11 bounces on the held-out one - [Supporting source](https://arxiv.org/abs/2610.38042) ## Source-stated The authors say RallyArb's hit timing was too far from the contact frame to match human timing. > it places hits too far from the contact frame to match human timing. - [Supporting source](https://arxiv.org/abs/2610.38042) ## Source-stated The vision-language models detected events from every viewpoint but missed the contact frame by about 0.1 s. > The vision-language models detect events on every viewpoint and miss the contact frame by about 0.1 s. - [Supporting source](https://arxiv.org/abs/2610.38042) ## Source-stated The authors report that fine-tuning the spotter reached about 0.8 bounce F1 on two phone and court-level clips but remained at 0.2 on a grazing phone angle. > Fine-tuned leave-one-clip-out from its broadcast weights, the spotter reaches about 0.8 bounce F1 on two phone and court-level clips and stays at 0.2 on a grazing phone angle. - [Supporting source](https://arxiv.org/abs/2610.38042) ## Source and contributor context Soeiro-Vuong, David, Rochet, Andre-Louis · Article author These are credits stated by the source; they do not establish a profile match or authorship of every passage. ## Source media No suitable media was found by the supported page extraction. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Benchmark A defined set of tests used to compare systems. Results depend on the tasks and measures selected. - [Definition reference](https://arxiv.org/abs/2211.09110) ## Fine-tuning Further training of an existing model to adapt it to a particular task or use. - [Definition reference](https://developers.google.com/machine-learning/glossary#fine-tuning) ## arXiv arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed. - [Explore this context](/entities/arxiv) - [Reference](https://arxiv.org/) - [Reference](https://info.arxiv.org/help/api/index.html) ## Editorial provenance Version 1. Publication ID: c31c793f-7ac6-4779-811d-57879f5d97b8. Prepared by gpt-6-luna; grounding review by gpt-6-luna. Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-13T04:18:25.924Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: material\_development. Observation: bc027260bdc6a9f4594fe6615e50f01d1f248eb9ec4d31ae090fd49bdbcf30ef. Source text hash: 0b36ee7926c0984f5787aeb16d607f79a56ee600663c6f9578ffbdb03bd9eaea. ### uncertain — 2026-11-13T04:18:25.924Z — anthropic/claude-opus-5.5 Publication: c31c793f-7ac6-4779-811d-57879f5d97b8 attribution — pass: The draft correctly credits David Soeiro-Vuong and Andre-Louis Rochet as the authors of the arXiv paper. It keeps results attributed to the authors' reported experiments rather than presenting them as independent fact. > Authors:David Soeiro-Vuong, Andre-Louis Rochet summary — uncertain: The numbers are faithful, but central terms that carry the meaning are unexplained for a non-specialist: F1, supervised spotters, published broadcast weights, arbitration layer, vision-language models and frozen predictions. Request tooltips for these. The draft also leaves out the stake the abstract gives: tennis statistics are derived from these events. Smallest repairs: say that once, add that RallyArb fits no parameter on event labels, note that the VLMs were zero-shot, and keep the leave-one-clip-out condition on fine-tuning. > Ball bounces and racket hits are the events from which tennis statistics are derived date — pass: originalPublished 2026-10-27 matches the v1 submission date. The later feed date of Oct 31 is correctly not used. > \[Submitted on 27 Oct 2026\] support — pass: Each claim matches the abstract, including the caveat that RallyArb was tuned on two of its three phone and court-level clips and the 8-of-11 held-out result. The closing line about a gap between recognizing events and timing them is a modest synthesis grounded in the VLM and RallyArb timing results, and generalization is left open. Adding the pooled E2E-Spot bounce F1 of 0.130 would strengthen the 'almost nothing' claim, but it is not required. > The vision-language models detect events on every viewpoint and miss the contact frame by about 0.1 s. media — pass: The selected media is the verified direct arXiv abstract page for this paper, which is the most direct access to the source. ## Recorded references Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter. No public references recorded here yet. [Reference data (JSON)](/readings/8143d01da73c9d42c04d356f22e7bfcf1ab81c4bed2f3ad99d463ef2f6578507/references.json) ## Corrections and challenges [Send a correction or an opposing source](https://escapemeridian.com/corrections/) [JSON record](https://escapemeridian.com/readings/soeiro-vuong-and-rochet-test-tennis-event-detectors-beyond-8143d01d.json)