ゼミナール発表

日時: 9月29日(火)4限 (15:10-16:40)


会場: L1

司会: 川波 弘道
川名 雄樹 1451039: M, 2回目発表 萩田 紀博,池田 和司,浮田 宗伯

title: *** Ensemble of Multiple Human Pose Estimation Models Using Convolutional Networks***

abstract: *** We propose a method to infer human pose in a single static image by ensembling multiple human pose estimation models. We first train individual pose estimation models with clustered training data (i.e. clustered by action being taken by a person in an image) to reduce variance of data in each cluster. Sequentially we train ensemble model which learns coocurrence (i.e. such as similarity of output on certain body parts between certain individual models ) among outputs of the pre-trained individual pose estimation models. We employ convolutional networks to exploit the coocurrence to optimize the final output of human pose and fine-tune each pose estimation model at the same time. ***

language of the presentation: *** Japanese***