roadscopeReplay lab
TIGER DATA DB 0 Potholes
RECORDED TEST DATA Export data
FROM SENSOR TO STREET

Drive replay.

Watch road conditions take shape, one observation at a time.

9 drives · Kaggle + LiRA
Consensus + hysteresis
LOADING DRIVELarisa, Greece
ESTIMATED ROUGHNESSGoodMediumBadTerribleDisturbanceTiger DB Pothole
Road colors appear after the prediction is finalized.
Preparing your driveLoading sensor readings and timed predictions…

Accelerometer m/s²

XYZ
Includes gravity100 Hz · last 12 s

Gyroscope rad/s

XYZ
Recorded sensor axes100 Hz · last 12 s

Model timeline

IRIDefect %
Solid: final · dashed: provisional160 ms patches
00:00.0Preparing replay00:00
—Every timestep · no windows skipped
ABOUT THIS REPLAY

A drive, seen through the model.

This is a real-time replay of recorded test data and precomputed predictions from the trained four-model ensemble. No new neural-network inference runs in your browser.

What appears when?

The car follows already observed GPS fixes. Predictions appear only after their exported availability timestamps. Final decisions combine overlapping contexts and arrive 320 ms after the target patch ends, plus any sensor interpolation wait. Hardware and network latency are not added.

Reading the map

Road colors show final estimated IRI using the project's bins: good below 2, medium 2–4, bad 4–6, and terrible from 6 m/km. Hollow dots are provisional. Coral markers show finalized disturbance onsets. GPS gaps over three seconds hide the current car; no future fixes fill gaps.

What the labels can tell us

The disturbance head detects a combined category of manholes, cracks, bumps and depressions. It is not a dedicated pothole classifier. Kaggle has disturbance labels but no measured IRI. LiRA has section IRI but no disturbance labels, and no gyroscope. The last two patches and any incomplete tail stay provisional.

Space: play/pause · ← / →: move 5 seconds · Use the timeline to explore any point in a drive.