Out-of-distribution

A benchmark focused on robustness under geographic shift.

Benchmark overview

The out-of-distribution benchmark measures robustness under geographic shift. Following the camera-ready protocol, the test set is restricted to U.S. footage while the remaining real data is available for training, and the split is kept comparable in size to the in-distribution setting so that performance differences are driven primarily by geographic shift rather than by changes in data quantity.

Tasks: temporal localization, spatial localization, and collision type classification.
Protocol: train on 454 clips and evaluate on a held-out U.S. test set of 1,573 clips.
Official score: unified ACCIDENT score together with the three task-specific metrics.

Score Your Submission

Select CSV File

Leaderboard

Ranked by unified score

This table tracks public results under regional shift, with method metadata rendered directly from submission records. Click a metric header to sort the ranking.

	Method	Paper

No scored submissions yet.