Researchers have introduced FPV-CRASH, a dataset built specifically for offline Crash Moment Localization in drone first-person-view videos. The task is to identify the single frame where physical impact interrupts controlled flight, giving a temporal reference for post-incident analysis. The authors note that existing crash-video datasets focus on surveillance or dash-camera footage, whose annotation schemes do not fit the onboard FPV setting.
The dataset comprises 1,052 clips: 526 crash and 526 normal. Each crash clip has a one-frame reference annotation for the crash moment, while normal clips provide background supervision during training. Every clip is paired with four English captions generated from sampled frames and the crash/normal label, supporting an auxiliary text-to-video retrieval task; localization itself remains video-only.
As a baseline, the paper presents a heatmap-based reference model that estimates a continuous crash timestamp from sparsely sampled frames. Under a fixed, non-source-disjoint clip-level protocol, it reduces mean absolute error from 1.12 seconds for the strongest fixed timing-prior baseline to 0.56 seconds, a 50% relative reduction, and raises the fraction of clips localized within 0.5 seconds from 37.14% to 65.71%. The release includes annotations, source video URLs, clipping metadata, fixed split lists, and preprocessing code, allowing the reportable portion to be reconstructed without redistributing source footage. At submission time, all source-compilation URLs remained accessible.