[Abstract]

 Supervised learning of Convolutional Neural Networks (CNNs) has driven recent advances in visual recognition. Due to the data-hungry nature of deep CNNs, this approach demands an enormous number of training images with groundtruth labels, which are given by hand in general. However, manual annotation of the labels is prohibitively time-consuming for high-level visual recognition tasks like semantic segmentation, which results in existing datasets limited in terms of both class diversity and the amount of labeled data. It is thus not straightforward to learn high-level visual recognition models that can handle diverse object classes in the real world. As a way to alleviate this issue, this talk suggests weakly supervised learning that adopts weaker yet less expensive and readily available labels as supervision. As well as the definition, motivations, and challenges of weakly supervised learning, recent research of our group on the topic will be introduced in this talk.

[발표자료]