Learning Hybrid Template by EM-Type Algorithm
Bin Lai, Dengyi Zhang, Chengzhang Qu, Jianhui Zhao, Zhiyong Yuan · 2009
The article proposes to improve the active basis model by incorporating both unaligned training examples and non-alignable sketches in images. EM-type algorithm [8] can learn the objects appear at unknown orientations, locations and scales in the training images. And non-alignable sketches [11] can be summarized in average sketches over the image lattice. This article proposes to add the score of the non-alignable sketches to the likelihood of M-step, so the learned active basis model by EM-type algorithm should be more accurate. Our experiments show that the proposed model can achieve considerable improvement in ROC for most of object categories.