CRxK dataset: a multi-view surveillance video dataset for re-enacted crimes in Korea.

We introduce a novel benchmark dataset, CRxK Dataset, featuring surveillance images based on meticulously re-enacted crime events with curated annotations. CRxK dataset includes a collection of crime re-enacted videos and images categorized into 13 different categories, encompassing a set of offenses such as assault, intoxication, swoon, and more. In addition, we cover a separate normal dataset extracted from scenes occurring five seconds before the crime event and spanning a 10-second duration. Among 13 categories, we employ six core categories, which are the most frequent occurrences in the collected dataset. These categories are assault, robbery, swooning, kidnapping, burglary, and normal scenarios. We conducted experiments using a total of 2,054,013 frames randomly selected and shuffled from the videos. Our training and validation involved four convolutional neural network (CNN) models and a single transformer model. We utilized a smaller sub-dataset, CRxK-6 dataset, containing 8,500 frames randomly sampled from each category video, resulting in 51,000 frames. Despite employing a train-test split ratio of 1:40 and applying face masking using RetinaFace, the dataset exhibited excellent performance with common CNN models, achieving an accuracy exceeding 0.940 for each model. However, it presented some challenges for the Transformer model.