The document proposes a method for detecting social events from photo metadata by applying watershed segmentation techniques to cluster images into events based on similarity of user annotations, with images assigned to the same event if they share common markers like username, date, tags or nearby GPS locations. The watershed transform treats each unique annotation as a marker and floods image regions from these markers using merging conditions to group images into homogeneous event segments, with boundaries between non-overlapping user involvement preventing images from being assigned to multiple events. External data sources on keywords, locations and semantic relatedness could help prune markers and improve precision and recall of identified social events.