COMBINING GLOBAL AND SIMPLIFIED PARTS-BASED APPROACH TO ESTIMATE HUMAN BODY CONFIGURATION(International Workshop on Advanced Image Technology 2005)

概要

論文の詳細を見る
We propose a novel approach to recover human body configuration from binary images. The basic idea is to search for the optimal configuration by Maximizing unified Likelihood Estimation (MULE). The unified-Likelihood is a linear combination of global (full body) and local (parts) likelihood. A simplified parts model of only upper body and lower body is used. The advantage of this model over other detailed parts models is sub-images corresponding to each part can be detected easily even in binary images. To avoid searching a wide configuration space, we narrow down the range of partial configurations for upper body as well as lower body, by retrieving candidate parts from training set based on parts-based likelihood where the input is two partitioned sub-images. These candidate parts are then combined exhaustively into a set of full body candidates, from among which we choose the optimal one based on the unified Likelihood. We conducted experiments with real images and results show good performance of the proposed unified-Likelihood and search strategy.
社団法人電子情報通信学会の論文
2005-01-03