作者: Xiaodan Liang , Shuicheng Yan , Jiashi Feng , Wen Feng Lu , Zequn Jie
DOI:
关键词: Machine learning 、 Pascal (programming language) 、 Traverse 、 Mathematics 、 Reinforcement learning 、 Perception 、 Interdependence 、 Feed forward 、 Data mining 、 Object-oriented design 、 Artificial intelligence
摘要: Existing object proposal algorithms usually search for possible regions over multiple locations and scales separately, which ignore the interdependency among different objects deviate from human perception procedure. To incorporate global between into localization, we propose an effective Tree-structured Reinforcement Learning (Tree-RL) approach to sequentially by fully exploiting both current observation historical paths. The Tree-RL learns searching policies through maximizing long-term reward that reflects localization accuracies all objects. Starting with taking entire image as a proposal, allows agent discover via tree-structured traversing scheme. Allowing near-optimal policies, offers more diversity in paths is able find single feedforward pass. Therefore, can better cover various quite appealing context of proposal. Experiments on PASCAL VOC 2007 2012 validate effectiveness Tree-RL, achieve comparable recalls much fewer candidate windows.