Video Dataset of Incidents (VIDI) is a general-disaster dataset for video incident classification. It contains 4,534 video clips of 43 incident categories and is introduced for multiclass classification.
ML task type: Multiclass classification
Data Source: Social Media (YouTube)
Size: 4,534 Videos
Timespan: N/A
Geographical Coverage: Global
Baseline Information
Evaluated on: Vision Transformer, TimeSformer
Metrics used: Top-1 accuracy, Top-5 accuracy
Results as reported in original paper: Top-1 accuracy: 76.54 (TimeSformer), Top-5 accuracy: 96.51 (TimeSformer)
Duygu Sesver, Alp Eren Gen¸co˘glu, C¸ a˘grı Emre Yıldız, Zehra G¨unindi, Faeze Habibi, Ziya Ata Yazıcı, and Hazım Kemal
Ekenel. VIDI: A video dataset of incidents. arXiv preprint arXiv:2205.13277, 2022.