Image annotation using personal calendars as context

In this paper, we introduce the idea of using the context of a personal calendar for labeling photo collections. Calendar event annotations are matched to images based on image capture time, and a Naïve Bayes model considers features from the calendar events as well as from computer vision-based image analysis to determine if the image actually matches the calendar event. This approach has the benefit that it requires no extra annotation from the consumer, since most people already keep calendars. In our test collections, 36 % of personal images could be tagged with a label from a personal calendar. Note that our preliminary results represent a lower bound on the performance that is possible because all of the system components are expected to improve over time. As people migrate toward digital calendars, we can also expect more consistency in their calendar labels, which should improve the annotation accuracy.