SECTOR

LIVE Monocular flight tracking · throws

One phone.One throw.
Every number.

SECTOR turns a single phone video of a discus, shot, hammer or javelin into how fast it left the hand, how steeply, how high and how far it truly went — each with an honest error bar.

  • Runs in your browser
  • Video never uploaded
  • No sensors or markers
●Flight solvedReconstruction · not broadcast footage
Release speed
25.3m/s
±0.45 · answer key 25.33
Release angle
35.9°
answer key 36.0
Release height
1.76m
±0.96 · answer key 1.75
Rule 32 distance
75.56m
±1.74 · answer key 75.56
Fig. 1Mykolas Alekna's 75.56 m world record (Ramona, 13 April 2025), rebuilt from the physics, rendered as video, then tracked with no access to the answer. The “answer key” is what the renderer actually threw.

§ 01 How it works

Four steps, from video to verdict.

No sensors, no markers, no second camera. The figures below are not mock-ups: they are the real output of the tracker on one rendered throw.

Step 01: Film it

Prop a phone on the ground, side-on, and record one throw. Keep it still and keep the circle and the whole flight in the picture.

Under the hoodAny phone video works. The camera must be static; the four edges of the ring are your only calibration. Decoded in-tab by video.ts.

frame 66 · 1.32 s
frame 110 · 2.20 s
frame 155 · 3.10 s
frame 196 · 3.92 s
Fig. 2Four of the clip's 238 frames. The ringed speck is the implement: a few pixels wide and smeared by motion. There is nothing to “recognise”, which is why SECTOR does not try.

Step 02: Find the throw

SECTOR looks for every speck that moved, in every frame. Most are wrong: people, lights, noise. It keeps only the specks that line up on a smooth arc.

Under the hoodMedian-background subtraction in detect.ts, then RANSAC in track.ts with a gate that rejects anything whose vertical acceleration is not gravity.

the arc that obeys gravity
● 149 frames kept● 523 specks thrown out
Fig. 3Every candidate from all 238 frames stacked on one picture (672 in total). Only one chain of them falls the way gravity says a thrown thing must.

Step 03: Solve the physics

A thrown thing obeys rules: gravity pulls it down and the air pushes back. Only one launch speed, angle and height can draw the arc you filmed. SECTOR finds it.

Under the hoodFour rim clicks give a homography and camera pose (solve.ts). An RK4 flight with lift and drag (physics.ts, aerofit.ts) is fitted to the arc.

04812150204060m from releaseheight, mreleaseapex 13.5 m+11.0 m from lift
measured flightsame release, no air
Fig. 4A discus is a wing, not a rock. Fit it with a plain parabola and you get 33.5 m/s and 115.4 m for this very clip: 8.1 m/s and 39.8 m too high.

Step 04: Report it, with error bars

You get release speed, angle, height and the legal distance. Each one says how far it could be off, so you know what to trust and what to ignore.

Under the hoodBars come from re-solving at ±8% field of view (solveWithUncertainty). Distance is measured the Rule 32 way, from the circle's inside edge.

Release speed25.33 m/s ±0.45
answer key 25.33 m/s · error 0.003
Release angle35.9°
answer key 36.0 ° · error 0.06
Release height1.76 m ±0.96
answer key 1.75 m · error 0.014
Rule 32 distance75.56 m ±1.74
answer key 75.56 m · error 0.002
Fig. 5The orange box is how far each answer moves if the phone's zoom is wrong by 8% either way — the honest error bar. The white tick is the answer key the renderer actually used.

§ 02 The Studio

Now you try it.

Watch SECTOR track a world-record throw, design a throw of your own, or drop in a phone video. It all runs in this tab.

  1. 1Pick a source
  2. 2Run the tracker
  3. 3Read the result

Pick a real mark. It is rebuilt from the physics, rendered as video, and re-measured by the tracker.

4.22s
Overlay
reprojection residual 0.20 px
25.3 m/s · 35.9° · 75.56 m

This result was solved when the page was built, with the same code. Press run to solve it again, live, in your browser.

blobs 672 @ threshold 11hypotheses 6 proposedgate 0 rejected by physicsarc 149 frames acceptedsolved at build time
Pick a mark
Reconstruction

The first man past 75 metres, thrown in wind at the venue that has now produced two world records. Drag the headwind and watch why athletes travel to Ramona.

VenueRamona, Oklahoma, USA
Date13 April 2025
Published mark75.56 m
Release required25.33 m/s
At angle36.0° · plate 28°
Lift is worth+11.0 m over vacuum
Venue headwind0.00 m/s

The release is fixed. Only the air moves — and a discus is a wing, so into a headwind it flies further. This is why athletes travel to Ramona. Then press run.

Venue produced75.56 m
SECTOR measured75.56 m

This is a reconstruction, not a measurement of the throw. SECTOR solves for the release this model needs to reach a published mark, renders that flight, then measures it back — so the error you see is the tracker's, not the athlete's. Source

Result: released at 25.3 metres per second, 35.9 degrees, official distance 75.56 metres.
Result
HIGH CONFIDENCEaerodynamic flight model · 0.20 px reprojection

It left the hand at 25.3 m/s (±0.45), 35.9° up, from 1.76 m high.
It went 75.56 m (±1.74) by the official measure. It landed right down the middle of the sector.

Release velocity
25.3m/s
answer key 25.33 · off by 0.00
Release angle
35.9°
answer key 36.00 · off by -0.06
Release height
1.76m
answer key 1.75 · off by +0.01
Rule 32 distance
75.56m
answer key 75.56 · off by 0.00
Flight time
4.00s
apex 13.5 m
Sector deviation
+0.0°
17.5° of margin to the line
Aero efficiency
1.171×
vacuum range 64.5 m
Plate attitude
28.0°
inferred from flight
What a parabola tracker would have told you
33.5 m/s · 115.4 m → off by +8.1 m/s and +39.8 m
A discus is an airfoil, not a projectile. Fit it with a parabola and the optimiser inflates release speed until the curve happens to match the pixels — with a respectable-looking residual and a badly wrong answer.
Same release, different wind
71
-4
73
-2
76
0
78
+2
81
+4
84
+6
Headwind is positive. A discus gains distance into a moderate headwind — what matters is airspeed over the plate, not groundspeed.

Reading the ± figures: they show how far each answer moves if the camera's field of view is off by 8% either way. Confidence reflects how well the arc fits, how clean the four rim points are, and how many frames the throw was seen in.

Flight profile
04812150204060m from releaseheight, mreleaseapex 13.5 m+11.0 m from lift
measured flightsame release, no air
Where it landed
102030405060708090CIRCLESECTOR 34.92°

§ 03 The archive

Real marks, measured back.

7 published throws, from a Jürgen Schult world record to Olympic gold. None of them was produced by this code, so none of them can be tuned to. For each, SECTOR asks one question: what release does the physics require to reach that exact distance? If the answer were 40 m/s, the model would be wrong. Every answer comes out at a believable release speed, between 22.8 and 25.3 m/s.

Release speeds are model output, not measurements of the athletes. Marks and sources are in src/lib/casefiles.ts; every one is checked to within a centimetre by casefiles.test.ts. Ruler runs 0–80 m. Broadcast footage is not redistributed: the archive ships marks and physics, never video.

§ 04 Evidence

Does it actually work?

On throws where the answer is known, yes — and the test suite fails if that ever stops being true. Each row below is a synthetic throw: the renderer knows exactly what it threw, SECTOR sees only the pixels, and the table is the gap between the two. These numbers were computed when this page was built, by the same code the tests run.

Measurement error against known ground truth, per test throw
Test throwTruthSpeed errAngle errHeight errDistance errReproj.
Reference throw68.14 m0.011 m/s0.092°0.013 m0.039 m0.20 px
Flat and slow41.08 m0.040 m/s0.071°0.012 m0.118 m0.21 px
Fast, steep, into 5 m/s83.39 m0.004 m/s0.044°0.015 m0.000 m0.21 px
Test fails above0.5 m/s1.5°0.15 m1 m2 px

Source: src/lib/accuracy.test.ts · reproduce with npm run accuracy · worst case uses 11.8% of its allowed error

A plain-parabola tracker would have been off by +8.1 m/s and +39.8 m on the same pixels.

A discus is a wing, so fitting it with a projectile's parabola does not fail loudly: it inflates the speed until the curve happens to match, with a respectable-looking residual. That is the case SECTOR is built around. (World-record reconstruction, from the demo above.)

Source: ThrowMetrics.ballisticComparison, same run as Fig. 1

Share of the allowed error each test throw actually used

Release speedlimit 0.5 m/s
2.2%
8.1%
0.8%
Release anglelimit 1.5 °
6.1%
4.8%
2.9%
Release heightlimit 0.15 m
8.7%
7.8%
10.2%
Distancelimit 1 m
3.9%
11.8%
0.0%
0%100% = test fails
Fig. 6One bar per test throw, in the order of the table. The bars are barely visible because the errors are tiny next to the limits — on purpose: the limits are set loose so only a real regression trips them.

What we have not proven.

A tool that only lists its wins is selling something. These are the edges of what the evidence above can claim.

  • 01

    Synthetic, not surveyed. The numbers above come from a renderer. They prove the detect, track, calibrate and solve chain is correct end to end. They do not yet prove the renderer matches a real phone camera. Field validation against taped marks is the next milestone; until then, treat real-footage output as a training aid.

  • 02

    Archive marks are reconstructions. They say what the physics requires to reach a published mark, not what any athlete's arm did.

  • 03

    Pinhole camera only. No lens-distortion model, so wide-angle footage biases the solve near the edges. Rolling shutter is not corrected either.

  • 04

    Static camera, visible ring. The background model assumes the camera never moves, and calibration needs the rim in frame.

  • 05

    Not an officiating device. It does not replace a steel tape or an official's call.