QMB Morning Dataset

The proliferation of wearable cameras has accelerated and facilitated a surge in research on analysis of egocentric videos. This relatively new field has relatively few research datasets. A significant share of publicly available egocentric datasets is purposely acquired for activity recognition, video summarization, object detection, and behavioural understanding. Here we share our dataset that is focused on personal localization and mapping.

We collected our dataset on the university campus, documenting a user’s typical morning. The recording is always initiated at the building entrance, at which point the user enters and triggers the session. The dataset was collected indoors before working hours to ensure people were not mistakenly captured, avoiding potential privacy challenges. To further guard ourselves, we recorded the dataset exclusively in an indoor environment to limit strangers’ appearance in the field of view. The five sessions cover a one-month period. During recording, the user’s operation was in line with a loosely worded script to ensure that multiple visits were made to a range of locations. In total, we captured nine distinct stations: Entrance, 3d-lab, Kitchen 1, Kitchen 2, Cafe-area, Lab, Printer 1, Printer 2 and Office.