Computer Vision News - July 2018

Yongqin Xian is a PhD student at Max Planck Institute for Informatics, with Bernt Schiele and Zeynep Akata . We spoke to him before his poster session on June 20 at CVPR. Yongqin tells us that in this work, they are trying to tackle the generalized zero-shot learning problem . They want to train their classifier to predict both seen and unseen classes at test time. This is different from conventional zero-shot learning, because in conventional zero-shot learning, the classifier only predicts unseen classes. This problem is very challenging, he says, because it suffers from extreme data imbalance between seen and unseen classes. There is a lot of seen classes data, but there is no unseen classes data at all. This makes the classifier bias to seen classes. That’s why in this work, he proposes to generate synthetic data of unseen classes to fix this data imbalance. However, instead of generating synthetic images of unseen classes like most GAN papers do, he proposes to directly generate synthetic CNN image features of unseen classes, conditioned on all class-level attributes. 38 Yongqin Xian Thursday From left: Yongqin Xian, Zeynep Akata and Tobias Lorenz

RkJQdWJsaXNoZXIy NTc3NzU=