报 告 人：张巍 博士 [复旦大学计算机学院 副教授, CCF优秀博士论文获得者]
报告时间：2013年12月17日（周二）13：00
报告地点：上海大学校本部计算机大楼901室
邀 请 人：李晓强 副研究员
报告简介：
Semantic segmentation is an interesting and challenging problem in computer vision, which aims to assign each superpixel in an image to one of pre-defined semantic categories. The key problem of semantic segmentation is to learn superpixel classifiers which identify semantic label for each superpixel in images. In traditional fully supervised methods, various machine learning techniques are used to train classifiers on labeled superpixels; however, in many practical applications, it is not easy to obtain enough labeled superpixels to learn satisfying classifiers for semantic segmentation. On the contrary, only image-level labels are necessary in weakly supervised semantic segmentation. We perform semantic segmentation in weak supervision by two approaches: I) We try to estimate the superpixel labels in the training set based on image-level labels such that superpixel classifiers can be trained. II) Alternatively, we select optimal classifier by parameters evaluation instead of training. More specifically, we firstly sample the classifier parameters at random and then evaluate the superpixel classifiers by measuring the reconstruction errors among the ground-truth negative samples and the predicted positive samples.
报告人简介：
张巍，男，副教授。1996年9月考入复旦大学开始本科学习，2008年1月获计算机应用技术专业的理学博士学位，2008年3月留校从事教学与科研工作。研究方向主要有机器学习、模式识别、图像内容分析与理解等。多篇研究论文发表在CCF推荐的A类学术会议上（如ICCV、CVPR、IJCAI、ICML等）。2008年度中国计算机学会优秀博士学位论文获得者。兼任Frontier of Computer Science 期刊青年AE。2013年1月入选复旦大学“卓学”计划。作为第一负责人，分别承担国家自然科学青年基金一项和承担上海市科委项目一项。
