Learning Visual Manifolds by Information Projection
讲座名称：
Learning Visual Manifolds by Information Projection
讲座时间：
2008-10-24
讲座人：
朱松纯
形式：
校区：
兴庆校区
实践学分：
讲座内容：
题目：Learning Visual Manifolds by Information Projection
报告人：朱松纯教授
时间：10月24日下午3：00-6：00
地点：科学馆324会议室
应西安交通大学校长郑南宁院士邀请，美国UCLA大学朱松纯教授将于10月24日至26日访问我校。朱松纯教授主要从事计算机视觉与模式识别，统计建模与学习，统计计算等方面研究工作。
主办：西安交通大学人工智能与机器人研究所
报告简介：
Images (and video) are very high dimensional signals that reside in a wide spectrum of manifolds of varying dimensions. But what are the structures of these manifolds? How are they related to each other in the universe of images? Traditionally, we have two mathematical tools for image modeling: (i) Markov random fields (MRF) from modern statistical physics; and (ii) sparse coding and wavelet from harmonic analysis. It is unknown in the literature how these two methods can be related to each other.
This presentation will illustrate:
1. two types of pure manifolds: (i) implicit manifold for high entropy patterns, like texture, modeled by MRF and (ii) explicit manifolds for low entropy patterns, like textons and image primitives, modeled by sparse coding.
2. a unifying theory for learning probabilistic models by manifold pursuit through information projection and the two theories (MRF and sparse coding) work in two image regimes (high and low entropy)
3. to discuss how these manifold are mixed to form middle entropy patterns and integrated to generate a primal sketch representation conjectured by David Marr in his influential book
4. to demonstrate how these manifolds are composed by a stochastic graph grammar to form object categories
5. ongoing work on video manifold learning and its potential applications in coding and tracking
相关视频