4D Cardiff Conversation Database (4D CCDb): A 4D database of natural,dyadic conversations

Abstract

The 4D Cardiff Conversation Database (4D CCDb) is the first
\n4D (3D Video) audio-visual database containing natural conversations
\nbetween pairs of people. This publicly available
\ndatabase contains 17 conversations which have been fully annotated
\nfor speaker and listener activity: conversational facial
\nexpressions, head motion, and verbal/non-verbal utterances. It
\ncan be accessed at http://www.cs.cf.ac.uk/CCDb.
\nIn this paper we describe the data collection and annotation
\nprocess. We also provide results of a baseline classification
\nexperiment distinguishing frontchannel from backchannel
\nsmiles, using 3D Active Appearance Models for feature extraction,
\npolynomial fitting for representing the data as 4D sequences,
\nand Support Vector Machines for classification. We
\nbelieve this expression-rich, audio-visual database of natural
\nconversations will make a useful contribution to the computer
\nvision, affective computing, and cognitive science communities
\nby providing raw data, features, annotations, and baseline comparisons.