OV-CAVED UCF-Crime: Annotations for Open Vocabulary Context Aware Video Event Detection

OV-CAVED UCF-Crime is a benchmark for Open Vocabulary Context Aware Video Event Detection (OV-CAVED), derived from the UCF-Crime video anomaly detection dataset. Status: this record currently contains only a README. The annotations will be uploaded as a new version of this record upon acceptance of the associated paper. The dataset extends conventional temporal anomaly annotations with the additional information required by the OV-CAVED framework. Each video is associated with a structured operational context describing the monitored scene and with natural-language event queries specifying the events to be verified. Queries are provided at three levels of granularity (coarse, mid, fine), together with plausible absent queries used as hard negatives, and are aligned with temporal supports, enabling query-conditioned evaluation at both row level and video level. The benchmark preserves the original UCF-Crime train/test partition, with a validation split held out from the training set. Training and validation annotations are generated automatically by the proposed annotation tool, while test-set queries and temporal supports are manually reviewed. For the test set, both the automatically generated and the human-refined operational contexts are provided. Instead of evaluating only whether a video segment is anomalous, OV-CAVED UCF-Crime allows models to be tested on whether a specific user-defined event is visually present under the operational conditions of the monitored scene. Note on third-party data: the dataset contains annotations only. The original videos are not redistributed and must be obtained from the official UCF-Crime release, under its terms of use. Dense video descriptions used as input to the annotation tool are taken from UCA-Crime. This resource accompanies the paper "Open Vocabulary Context Aware Video Event Detection" by V. Carletti, A. Greco, M. Marseglia and M. Vento (University of Salerno), submitted to the International Journal of Computer Vision. Code: https://github.com/MiviaLab/ovcaved