Reporters: R Yun-Nung Chen, R Yu-Cheng Liu
Ming Actor Correlations with Hierarchical Concurrence Parsing (ICASSP 2010) Kun Yuan, Hongxun Yao, Rongrong Ji, Xiaoshuai Sun Computer Science & Technology, Harbin Institute of Technology
Introduction Actor Indexing Mining Actor Correlations Context-Based Actor Concurrence Graph Ranking Concurrent Shots Actor Correlation Changes Analysis Experimental Results
Actor correlations graph interfaces
Top 20 shots
Shot boundary detection (SBD) Shots and scenes Locating actor faces & face tracking algorithm Face set: different poses from the same actor
2D-PCA reduces dimension Features of same person may distribute discretely in feature space Given 2 face sets F k and F l, 2 pose sets If distance < T, 2 face sets belong to the same person
A shot and its surrounding shots may present a plot between two actors in video Gaussian weight measurement
Scene level correlation Video level correlation
Construct correlations graph from
A single character i Character correlations between i and j
Given i, j, sort RankScore (k) for all k Show top 20 shots
Two actors’ correlation changes with story Analyze the difference of concurrence R(i, j) A correlation measure between i and j in the part A Change ratio Hlp
20 hours video of “Friends” TV series About 4000 shots Over 800 face sets Clustering into about 60 face sets (T = 0.25) Manual labeling to 17 actors
The actor concurrence precision in all ranking shots is up to 90% The precision of each two actor’s co-occurrence in ranking top 20 is up to 98%