Proceedings of the 2nd ACM International Workshop on Human-centered Event Understanding from Multimedia

Abstract

It is our great pleasure to welcome you to the 2nd ACM International Workshop on Humancentered Event Understanding from Multimedia -- HuEvent'15. This year's workshop continues its tradition of being the premier ACM workshop for presentation of research results and experience reports in the field of events in multimedia. The mission of the workshop is to present and discuss the different aspects and notions of events and objects. This includes methods for detecting activities and low-level events and objects from media content and other sensory data. It also targets solutions and approaches for detecting and modeling the relationships between events and objects. The keynote is held by Cees Snoek on the topic of "Recognizing events in videos without examples". The talk presents recent progress on recognizing events in videos, without the need for examples. The key to event recognition in such a challenging setting is to have a lingual video representation. Three lingual representations for zero-example event recognition are highlighted: covering concept, tag, and sentence embedding. The accepted research papers at the workshop are continuing the discussions focusing on the following topics: The first paper, "Using Photo Similarity and Weighted Graphs for the Temporal Synchronization of Event-Centered Multi-User Photo Collections", addresses the issue of temporal synchronization of photo collections that have been created during the same event by different users using different (unsynchronized) devices. The method proposed by the authors employ multiple similarity measures to identify pairs of similar photos and then temporally align the photo collections by traversing a graph, whose nodes represent the collections, and edges represent the similar photo pairs between collections. The second paper, "Media Synchronization and Sub-Event Detection in Multi-User Image Collections", continues the idea of media synchronization by conducting a thorough evaluation of the performance of several visual-based image synchronization techniques; but goes beyond and applies it to sub-event detection. Common clustering techniques are experimented for the detection of sub-events in the presence of synchronization misalignment. Finally, the paper "Discovering Commonness and Specificness for Human Action Recognition" brings into discussion another dimension of human-centered event understanding, namely human action recognition. A discriminative dictionary, learning-based method that is specifically adapted to recognize commonness and specificness among different action classes is introduced. Experimental validation on standard benchmarking datasets show promising results for this approach.