TTennis reconstruction active research system

Broadcast tennis computer vision

Video → point ledger → event anchors → rally geometry.

The current system segments matches, tracks score, follows the ball in image space, and localizes racket hits, bounces, and net interactions. The immediate work is making those 2D anchors exact enough to support 3D reconstruction.

Execution graph2026-07-24
Full matchvideo + audio
Point ledgerwindows + score
2D trackframe observations
Anchorshit · bounce · net
3D rallyball + players
Every stage emits confidence.Every uncertain case reaches the portal.

System state

What is built, and what each stage hands forward.

The pipeline is the research object. Intermediate visualizations are part of the contract, not presentation added after the fact.

StageInputOutputState
01

Point ledger

All downstream work is point-addressable and reviewable in the portal.

Full-match videoPoint windows + score stateOperational
02

2D ball track

The decoded track is treated as the measurement substrate, not something physics may overwrite.

Native-frame cropsDense image-space trajectoryOperational
03

Event anchors

Each event carries frame, image position, confidence, witnesses, and an abstain path.

Track discontinuities + players + audioHit / bounce / net / artifactActive audit
04

3D reconstruction

Existing physics and factor-graph prototypes are retained as tools, not promoted as the current system.

2D rays + court planes + player poseConnected rally geometryNext

Current evidence

Scrub the measurements, not a polished reconstruction.

This replay uses the exact decoded coordinates and current event outputs from pt0143. The public site omits copyrighted broadcast pixels; the private audit portal places these same overlays on the source frames.

decoded ball trackimage coordinates · 960 × 540
frame 194x 721.9 · y 358.1
f194

Fresh audit excerpt

Sub-frame predictions beside reviewer-selected frames.

Round three exposed one between-point false proposal, two one-frame placement gaps, and an occluded contact. The current revision addresses all four mechanisms.

PointPredicted frameTypeConfidencePrimary witnessHuman frame
pt0117756.50racket_hit0.80horizontal reversal756
pt0117815.10bounce0.80ground vertical reversal815
pt0143201.68racket_hit0.80trajectory reversal202
pt0143457.50bounce0.80ground vertical reversal457
pt0143497.56racket_hit0.85wrist reach + occluded redirect498
Scope noteThese are reviewed development examples, not a broad generalization result. The next evaluation expands points, players, courts, and broadcast conditions only after the anchor flow is stable.
contact_audit_flow.html · pt0143 · f498
decoded trailhuman adjustment

Website-first method

The review interface is part of the detector.

01

Propose

The pipeline emits a typed event, sub-frame timing, location, confidence, and witnesses.

02

Inspect

The reviewer scrubs the original point with track history and competing overlays visible.

03

Resolve

Confirm, adjust, reject, or add a replacement event. Every action is locally recoverable.

04

Regress

The exported labels become a frozen test and the same page is regenerated after each revision.

Promotion sequence

The order is constrained by geometry.

Smooth-looking 3D is not useful if contact timing is wrong. Each promotion depends on the stage before it surviving both quantitative checks and direct visual inspection.

  1. Now

    Close 2D anchors

    Resolve remaining timing and classification failures in the audit flow.

  2. Next

    Broaden evaluation

    Freeze more points, then add new players and broadcast geometries.

  3. Then

    Lift ball + players

    Use calibrated rays, court planes, pose, and exact events to reconstruct rallies in 3D.

  4. After

    Estimate speed + spin

    Fit physical parameters only after the spatial chain is stable and reviewable.