MM-Store Dataset: multi-view and multi-modal dataset in a store environment

This dataset is currently not publicly available, but it is scheduled to be released as soon as possible in early 2026. MM-Store are datasets of human behavior in convenience stores recorded using distributed multimedia sensors. This dataset was recorded in the convenience store shown in Figure 1. Twelve omni-directional microphones and six cameras were used to record audio and video. The amount of data was 990 clips per location and sensor, divided into 792, 66, and 132 clips each for training, validating, and testing. Each clip, ranging from 30 to 90 seconds, contained events selected from 19 classes of events. Each event is designed to simulate an action in a convenience store and is performed by one clerk and one customer each. The labels for training were given as multi-labels indicating what events each clip contained. For evaluation, only the test data was given a strong label containing the onset/offset time of each event.