Organizations: School of Communication Engineering, Hangzhou Dianzi University, China. · School of Computer Science, Faculty of Science, University of Auckland, Auckland, New Zealand. · Department of Computer and Information Science, Faculty of Science and Technology, University of Macau, Macau, China.
Crash Moment Localization aims to localize the major physical impact that interrupts controlled flight in an onboard first-person-view (FPV) video recorded by a drone, thereby providing a temporal reference for retrospective crash analysis in aerial robotics. However, existing crash-video datasets predominantly target surveillance or dash-camera footage, and their annotation schemes and task formulations are not designed for temporal localization of the major physical impact in UAV onboard FPV videos. To address this gap, we present FPV-CRASH, a dataset and fixed-protocol reference evaluation for offline Crash Moment Localization in drone FPV videos. FPV-CRASH contains 1,052 drone FPV clips (526 Crash and 526 Normal); each Crash clip has a single-frame reference annotation for the Crash Moment, and every clip is paired with four English captions automatically generated from sampled frames and the known Crash/Normal label. Normal clips carry no timestamp target and provide training-time background supervision. The captions support an auxiliary Text-to-Video Retrieval task, whereas localization inference remains video-only. We further provide a heatmap-based reference model that estimates a continuous Crash Moment timestamp from sparsely sampled frames. Under the fixed, non-source-disjoint clip-level protocol, the reference model reduces MAE from 1.12 s for the strongest fixed timing-prior baseline to 0.56 s (a 50% relative reduction) and achieves a median absolute error (MedAE) of 0.33 s; the fraction of clips localized within 0.5 s rises from 37.14% to 65.71%. The release includes annotations, available source video identifiers and URLs, clipping metadata, fixed split lists, and preprocessing code so that the reportable portion can be reconstructed without redistributing source footage. At submission time, all source-compilation URLs remained accessible.
Figures & tables
Fig. 1: Overview of FPV-CRASH and Crash Moment Localization. The dataset contains Crash and Normal FPV videos with Crash Moment annotations. Given a fully observed FPV video, the task is to localize the major physical impact that interrupts the flight.
Fig. 2: Task definition and dataset statistics of FPV-CRASH. Red curves are Crash responses; green dashed and blue solid lines denote reference and predicted moments. Normal clips have no Crash Moment; text strips show truncated captions. Right panels show Crash/Normal clip-duration and Crash Moment distributions in 2-s bins; the lower solid curve is smoothed and the dashed line marks the 7.0-s median.
Partition
Crash
Normal
Clips
Captions
Train
316
316
632
2,528
Validation
105
105
210
840
Test
105
105
210
840
Total
526
526
1,052
4,208
TABLE I: FPV-CRASH dataset composition and fixed splits.
Fig. 3: The proposed heatmap-based localization framework with auxiliary video–text alignment. Temporal video features predict a Crash heatmap and continuous timestamp; text-video alignment is used only during training, while localization inference is video-only. The shared embedding space also supports Text-to-Video Retrieval.
Fig. 4: Localization results on the 105 Crash clips in the fixed test split. Top: SigLIP2 Full predictions versus reference Crash Moment timestamps; the dashed line denotes exact prediction and the shaded band indicates a ±1 -s tolerance. Bottom: empirical absolute-error CDFs for SigLIP2 Full and the train-relative-median prior; vertical dotted lines mark 0.5-, 1-, and 2-s tolerances.
TABLE IV: Backbone comparison under Crash-and-Normal training. All models use PE; Text denotes training-time text supervision through video–text alignment. Errors are in seconds.
TABLE VI: SigLIP2 Crash-only 40-epoch control (760 updates), approximately step-matched to 20-epoch Crash-and-Normal training (780 updates). Errors are in seconds.
TABLE VII: Loss ablations for SigLIP2 Full with Crash-and-Normal training. The w/o all heatmap BCE variant removes both Crash and Normal heatmap terms; other variants remove the named component, and w/o alignment uses the PE-only configuration from Table V .
T2V
V2T
Backbone
R@1
R@5
R@10
R@1
R@5
R@10
EVA01
35.71
72.86
88.10
37.14
79.52
89.05
SigLIP2
41.90
80.00
88.57
38.57
78.10
88.57
PE-Core
35.24
76.67
87.62
35.71
79.05
88.10
TABLE VIII: Auxiliary bidirectional retrieval for Full models under Crash-and-Normal training. Each direction contains 210 queries and 210 candidates. Values are Recall@K (%).