arthur.shhome

Predicting roulette
from video.

Roulette is deterministic, but small differences at the start of a spin compound quickly.

The project recovers that motion from video, then tests how far ahead it can be forecast.

Read the paper
  1. 01VIDEO42 source videos
  2. 02LABELSsegmentation data
  3. 03TRACKball and wheel
  4. 04NORMALISEcamera geometry
  5. 05PREDICTthree LSTM tasks
  6. 06TESTheld-out results
DATASET EXPLORER / 01ACQUIRE

TRACKING EXAMPLES

Select an image to see what the detector recorded.

Recovering
the motion.

Each camera view has to become a clean, comparable trajectory before it can be used for prediction.

Source footage

The dataset contains 42 public videos and 5,463,775 frames, recorded across different wheels, lighting conditions, and camera angles.
42
videos
5,463,775
frames
1.36 px
centroid error
PIPELINE / 01INPUT
DETECTED ELLIPSEHTOP-DOWN UNIT CIRCLE
ANGLE θ (rad)RADIUS r (unit)t*0s5s10s
TRACKSMOOTHDERIVEθ + r + θ̇ + ṙ + t*
PULL DISTANCE0.0px
VECTOR SPEED260px/s
STATEREADY · READY
0321519421225173462713361130823105241633120143192218297281235326SLOPED INWARD ↘OCCLUSION ZONEGRAB · TAP · FLING
RADIAL SLOPE (Δr / frame)MONITORING
+.0060−.0060s5s10s
OBSERVATION HORIZON
MODEL DEVELOPMENT / 01SPLIT

2,765 VALID SPINS

TRAIN
75%
VALIDATE
12.5%
TEST
12.5%
INPUTT × 11
features
BALL TRAJECTORY
LSTM128 / 128 units
OUTPUTsin θ̂, cos θ̂
RESTORE BESTepoch 0early stop
training lossvalidation loss
OPTIMISER
Adam
BATCH
64
TRAJECTORY LR
5e−4
WEIGHT DECAY
1e−5
EARLY STOPPING
10 epochs

More context,
lower error.

Across all three tasks, predictions improved as the observed window grew.

INPUTTIME MAEBALL MAEWHEEL MAE
MAE (frames)MAE (rad)59.190.8450.10810044.880.3490.03920037.050.1850.03430029.820.1240.03340025.940.1170.025500■ DROP-OFF TIME□ BALL TRAJECTORY▫ WHEEL TRAJECTORYAT T = 50010.0sobserved0.52stiming MAE0.117ball MAE (rad)0.025wheel MAE (rad)

BASELINE COMPARISON AT T = 300

RANDOM1.572
LAST-ANGLE1.565
LSTM0.185

At six seconds of input, the ball-path model is compared with two simple baselines.

WHERE IT FAILS

Sudden slowing and missing frames still cause large errors.
OCCLUSIONstable motionabrupt decelerationerror tail

Read the full paper.

The dissertation contains the complete method, experiments, limitations, and results.

Read the dissertation