ROVI-SPORTS · USER MANUAL

ROVI-Sports User Manual

Set-up, calibration, shot capture, processing, reading results and GSPro connection for the evaluation version of ROVI-Sports, a golf launch monitor that films the ball and club with two high-speed cameras and calculates shot data.

1. Introduction

ROVI-Sports is a golf launch monitor. Two high-speed monochrome cameras (820 fps) standing beside the golfer film the ball and clubhead at impact; AI finds and tracks the ball and club and calculates the shot data.

Cam1 (top, wide view) and Cam2 (bottom, close-up of impact) filming driver, 7-iron and sand wedge shots at 820 fps — about 137 times slower than real time

What it measures

GroupMetrics
Ball (7)Ball speed · launch angle (LA) · launch direction (LD) · total spin · spin axis · back spin · side spin
Club (5)Club speed · club path · attack angle · face angle · dynamic loft

Workflow

StepTaskChapter
1Position the cameras, set exposure and lighting3
2Calibration capture (once after set-up, again whenever a camera moves)4
3Capture shots5
4Process → check results · send to GSPro6 · 7 · 8

2. Components and specifications

Components

ItemDetails
CamerasTwo high-speed monochrome cameras (820 fps, global shutter) — Cam1 wide view, Cam2 close-up of impact
Lighting850 nm infrared lighting (or diffused LED)
MountA vertical rail holding the two cameras one above the other
Processing PC (evaluation version)Ubuntu 22.04 · Python 3.10 · NVIDIA GPU
Product (in development)Dedicated NVIDIA Jetson device + Companion App

Recommended processing PC (evaluation version)

ItemMinimumRecommended
CPU4 cores, 2.5 GHz8 cores, 3.0 GHz or more
GPUNone (runs on CPU, slowly)NVIDIA RTX 3060 6 GB or better
Memory8 GB16 GB or more
Storage10 GB (software and models)500 GB (to keep captured data)
For measurement accuracy, see the product page and the evaluation report provided by FLOYD.

3. Set-up

3.1 Positioning the cameras

Stand the camera rail beside the ball's target line, on the side opposite the golfer. The swing arc and follow-through pass on the golfer's side and toward the target, so even 30 cm from the ball the club does not hit the cameras.

  1. Stand the rail vertically (check with a level). The bottom end of the rail is about 300 mm from the ball.
  2. Cam2 (bottom): mount it at the ball end of the rail, facing the ball. About 300 mm to the ball (not closer than the lens's 150 mm minimum focus distance).
  3. Cam1 (top): mount it at the other end of the rail and tilt it slightly up so that it sees the ball and the space above it where the ball flies. About 580 mm to the ball (50–70 cm between the cameras).
  4. Check that the ball is about 146 px across in Cam2 (at 300 mm). Below 75 px, spin is hard to measure; move Cam2 closer in 50 mm steps.
You do not need to measure positions to the millimetre. Calibration (chapter 4) works out the exact positions.

3.2 Exposure and lighting

ItemSetting
Exposure time100 µs or less (prevents blur)
Auto exposureOff (same brightness every shot)
GainAs high as possible without noise
Ball area brightnessMean 130–180 (on a 0–255 scale)
Saturation (blown-out pixels)Under 5%
LightingDiffused light from 45° on both left and right, within 500 mm of the ball. Avoid strong reflections on the ball

3.3 Software

A PC installed by FLOYD is ready to use. On a new PC, unpack the FLOYD evaluation package and run:

pip install -r requirements.txt
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121   # NVIDIA GPU (CUDA 12)
python3 -c "import torch; print(torch.cuda.is_available())"                       # True means the GPU is used

4. Calibration

Calibration works out each camera's lens properties, where the cameras are relative to each other, and the ground and target direction (world coordinates). It has the biggest effect on accuracy, so always do it after set-up and again whenever a camera moves. Finish it before capturing shots.

CaptureContentsAmount
A. Chessboard pairsCam1 and Cam2 photograph the same chessboard at the same time30 pairs or more, spread evenly over 9 areas of the picture
B. Marker boardOne simultaneous Cam1 + Cam2 pair of a marker board on the groundAll 6 markers visible in both pictures

4.1 Chessboard

ItemValue
Pattern11 × 16 inner corners (12 × 17 squares)
Square size23 mm (for another size, enter the real size when calculating)
MaterialA flat board that does not bend, printed to within 0.1 mm

4.2 Marker board

Place six markers on a board lying level on the ground, at the positions below. The direction from the origin zero toward pos10_x becomes the target direction.

MarkerPosition (X, Y, Z) m
zero(0.00, 0.00, 0) — origin
pos10_x(0.10, 0.00, 0)
pos20_x(0.20, 0.00, 0)
pos30_x(0.30, 0.00, 0)
pos10_y(0.00, 0.10, 0)
pos5_x_pos5_y(0.05, 0.05, 0)

Right after the chessboard shots, without moving the cameras, take one Cam1 + Cam2 pair and save it in calImage_Marks/.

4.3 Calculation

python3 tools/calibration/calibrate_rig_and_world.py --config <calibration config file> --square_size_mm 23

The result (rig and world coordinate file) is linked into the processing settings. If this is your first time writing the config file, send the captured calImage/ and calImage_Marks/ folders to FLOYD and we will calculate it for you.

5. Capturing shots

5.1 Capture settings

ItemSetting
Frame rate820 fps (fixed)
SyncCam1 and Cam2 capture together, triggered at impact
Frames15 per shot — 1–2 before impact and 8–10 or more after impact
Ball placementThe same spot on the tee (within ±5 px on screen)
BallBoth plain and marker balls are processed (spin is steadier with marker balls)

5.2 Folder layout

Make a folder for each shot under a folder for each club, and put the 15 frames from both cameras in it. The club folder name (Driver, 7Iron, SW) selects the club recognition model, so keep these names.

data/
├── Driver/
│   ├── 1/
│   │   ├── Cam_1_roi_1.bmp … Cam_1_roi_15.bmp
│   │   └── Cam_2_roi_1.bmp … Cam_2_roi_15.bmp
│   ├── 2/
│   └── …
├── 7Iron/
└── SW/

5.3 Checking capture quality

For the first few shots, open the Cam2 frames and check:

CheckTargetIf not
Ball diameter75 px or more (about 146 px recommended)Move Cam2 toward the ball in 50 mm steps
BrightnessBall area mean 80–190 (130–180 recommended)If dark, increase exposure or lighting; if bright, shorten exposure
SaturationUnder 5%Cut exposure by 30–50% or move the lights
ContrastDimples and markers are sharpAdd diffused side light, remove direct light
Frames after impact8 or more frames with the ball visibleTrigger earlier, or move Cam2 a little further away (up to 400 mm)

6. Processing

./run.sh single data/Driver/1 data/Driver/1 out/Driver/1     # one shot: <Cam1 folder> <Cam2 folder> <output folder>
./run.sh batch  data/ out_batch/                             # every club and shot under data/

When both cameras' frames are in one folder, give the same folder for Cam1 and Cam2.

Main options when running directly:

python3 src/golf_tracker_server.py --config configs/config_multiview_v1_default_sign_fixed.yaml \
    --cam1 data/Driver/1 --cam2 data/Driver/1 --output out/Driver/1 --artifact_mode production
OptionContents
--artifact_mode productionSave result files only (the default diagnostic also saves check images and intermediate data)
--data_dir / --output_dirMany shots at once (same as batch)
--gsproSend results to GSPro (chapter 8)
--disable_club_poseTurn off club keypoint recognition (ball data only)
An NVIDIA GPU is recommended for processing. It also runs on the CPU alone, but much more slowly.

7. Reading results

Each shot's results are saved in its output folder.

FileContents
roi/roi_result.csvOne line of shot data (table below)
roi/shot_metrics.jsonAll ball, club and spin results with confidence
trajectory.csvThe ball's 3D trajectory
club_metrics.jsonClub results and recognition status
launch_conditions.yamlLaunch conditions summary

roi_result.csv columns

ColumnMeaningUnit
Ball speedBall speedm/s
Launch DirectionLaunch direction (left–right)°
Launch AngleLaunch angle (up–down)°
Total spinTotal spinrpm
Spin axisSpin axis tilt°
BackSpin / SideSpinBack spin and side spin componentsrpm
Club SpeedClubhead speedm/s
Club pathClub path (swing direction)°
Attack AngleAttack angle (up–down angle of approach)°
Face AngleFace angle°
Dynamic LoftDynamic loft°

Example (driver shot):

shot ID,Ball speed,Launch Direction,Launch Angle,Total spin,Spin axis,BackSpin,SideSpin,Club Speed,Club path,Attack Angle,Face Angle,Dynamic Loft
roi,56.92,-4.43,6.72,3941.2,-3.63,3933.3,-249.4,36.15,-5.75,-5.85,-9.49,10.91
The signs of the angles (left–right, up–down) follow the world coordinates set by the marker board in 4.2 (origin and target direction). If another device uses a different sign convention when you compare values, contact FLOYD.

Club data comes from keypoints on the clubhead (shaft, heel, toe, top and bottom of the face, and so on) found in Cam1 and Cam2 and calculated in 3D.

Club keypoint recognition — Cam1 (top) and Cam2 (bottom), 15 frames around impact

8. GSPro connection

Shot results can be sent to the golf simulator GSPro.

  1. In GSPro, turn on UDP under Settings → Connect → External Devices (default port 921).
  2. Add --gspro when processing.
python3 src/golf_tracker_server.py --config configs/config_multiview_v1_default_sign_fixed.yaml \
    --data_dir data/ --output_dir out_batch/ --gspro --gspro_host 127.0.0.1 --gspro_port 921
OptionContentsDefault
--gspro_hostAddress of the PC running GSPro127.0.0.1 (same PC)
--gspro_portGSPro port921
--gspro_speed_unitBall speed unit (mph / ms)mph

9. Companion App (UX preview)

The design (UX concept) of the phone app that will be used with the dedicated device in the product version. The current evaluation version (chapters 1–8) has no app, and screens and features may change during development. Figures on the screens are examples.

Companion App UX concept — the nine screens from onboarding to course simulation, played automatically

Screens

ScreenContents
01 OnboardingProduct intro — ball flight, spin and club delivery from three angles
02 Sign inAccount sign-in and linking
03 PairFinds a nearby launch monitor and connects (shows signal strength)
04 SessionClub choice, the "Hit when ready" screen, the last three shots with grades, video + data recording, end session
05 Shot resultSpin (backspin, sidespin, axis), top view (face to path, club path, club face, launch direction), side view (attack angle, launch angle), carry, ball speed, smash. Scroll down for the AI read and drill suggestions
06 3D trajectoryEvery shot of the session in 3D — iso, top and side views, drag to orbit, suspect shots marked
07 HistoryMonthly average carry, best and shot count, carry trend, session list
08 CompareTwo shots side by side, comparing trajectories and figures
09 CoursePick a course and simulate a round
Session, shot result and 3D trajectory screens (UX concept, example figures)
Session, shot result and 3D trajectory screens (UX concept, example figures)

Display modes and units

10. Troubleshooting

SymptomWhat to check
The ball is not foundIs the Cam2 ball area bright enough (3.2), is the ball in the set spot (5.1)?
Club keypoints are not foundIs the clubhead visible in the Cam2 impact frames (around the 4th–5th), is the club folder named correctly (5.2)?
Spin values jump aroundBall diameter, contrast and saturation (5.3); marker balls are recommended
Launch angle or direction leans to one sideCalibrate again (chapter 4) — a camera may have moved slightly
Processing is slowDoes python3 -c "import torch; print(torch.cuda.is_available())" print True; use --artifact_mode production
Shots do not reach GSProGSPro UDP on (chapter 8), port 921 allowed through the firewall, same network

If a problem continues, send us the shot folder (both sets of 15 frames) together with the output folder.

11. Contact

CompanyFLOYD Inc.
Emailfloyd@floyd.kr
Phone+82-70-8676-7976
Webfloyd.kr — Review request

For evaluation units or OEM applications, send us a review request and we will get back to you.