FPV-CRASH
A dataset and benchmark for localizing crash moments in first-person view (FPV) drone videos. It contains 1,052 clips labeled as Crash or Normal, with single-frame reference annotations provided for Crash clips. It also includes an auxiliary text-to-video retrieval task.
Timeline (last 90 days)
FPV-CRASH is a dataset and fixed-protocol evaluation benchmark for localizing the moment of a crash in offline footage, designed specifically for drone onboard first-person-view (FPV) video.
The task goal of Crash Moment Localization is to localize, in drone onboard FPV video, the moment of the main physical impact that interrupts controlled flight.
Crash Moment Localization requires single-frame-level timestamp regression rather than clip-level binary classification; Crash clips must provide a unique reference-frame annotation, while Normal clips carry no timestamp target.
During inference, FPV-CRASH maintains purely video input and does not rely on any text prompts.
Under the fixed non-source-separated clip-level protocol, the FPV-CRASH reference model achieves MAE 0.56 seconds, MedAE 0.33 seconds, and a localization-within-0.5-seconds rate of 65.71%, while the strongest fixed temporal prior baseline has an MAE of 1.12 seconds and a localization-within-0.5-seconds rate of 37.14%.
The above FPV-CRASH benchmark results use a non-source-separated protocol, in which the same source video may appear in both the training and test splits, so the numbers should be regarded as an upper-bound reference for the reference implementation rather than proof of cross-domain generalization ability.
The FPV-CRASH release includes annotation files, available source video identifiers and URLs, clip metadata, fixed split lists, and preprocessing code, but does not redistribute the original video material.
According to the paper, all compiled source video URLs were still accessible at the time of submission.
Crash videos are naturally scarce and difficult to annotate; real crash footage often comes from individual submissions by pilots, with severe viewpoint shake and motion blur, and the impact moment may last only one or two frames.
All Facts (9)
Crash videos are naturally scarce and difficult to annotate; real crash footage often comes from individual submissions by pilots, with severe viewpoint shake and motion blur, and the impact moment may last only one or two frames.
50%UnverifiedThe task goal of Crash Moment Localization is to localize, in drone onboard FPV video, the moment of the main physical impact that interrupts controlled flight.
50%UnverifiedCrash Moment Localization requires single-frame-level timestamp regression rather than clip-level binary classification; Crash clips must provide a unique reference-frame annotation, while Normal clips carry no timestamp target.
50%UnverifiedDuring inference, FPV-CRASH maintains purely video input and does not rely on any text prompts.
50%UnverifiedUnder the fixed non-source-separated clip-level protocol, the FPV-CRASH reference model achieves MAE 0.56 seconds, MedAE 0.33 seconds, and a localization-within-0.5-seconds rate of 65.71%, while the strongest fixed temporal prior baseline has an MAE of 1.12 seconds and a localization-within-0.5-seconds rate of 37.14%.
50%UnverifiedThe above FPV-CRASH benchmark results use a non-source-separated protocol, in which the same source video may appear in both the training and test splits, so the numbers should be regarded as an upper-bound reference for the reference implementation rather than proof of cross-domain generalization ability.
50%UnverifiedAccording to the paper, all compiled source video URLs were still accessible at the time of submission.
50%UnverifiedFPV-CRASH is a dataset and fixed-protocol evaluation benchmark for localizing the moment of a crash in offline footage, designed specifically for drone onboard first-person-view (FPV) video.
50%UnverifiedThe FPV-CRASH release includes annotation files, available source video identifiers and URLs, clip metadata, fixed split lists, and preprocessing code, but does not redistribute the original video material.
50%