Zero-shot

A benchmark for methods that operate without benchmark-specific labeled training data.

Benchmark overview

The zero-shot benchmark targets settings in which no labeled training data is available. Methods are evaluated directly on the real dataset without benchmark-specific supervision, reflecting deployment scenarios where accident-specific labels cannot be collected in advance and favoring prompt-based, instruction-based, and other training-free approaches.

Tasks: temporal localization, spatial localization, and collision type classification.
Protocol: no benchmark-specific labeled training partition is used; evaluation is performed on all 2,027 real clips.
Official score: unified ACCIDENT score with the three task-specific metrics reported separately.

Score Your Submission

Select CSV File

Leaderboard

Ranked by unified score

This table tracks methods evaluated without benchmark-specific supervision and links results back to their paper or code when provided. Click a metric header to sort the ranking.

	Method	Paper

No scored submissions yet.