Speaker: Yizhou Wang Title: Statistical Inference on Markov Random Graphs Time: Friday(March, 18) 12:300-1:30pm Room: BH4750 Abstract As natural scenes contain a huge number of visual patterns generated by various stochastic processes, how to represent and model these diverse visual patterns, and how to learn and compute/infer those visual patterns efficiently become fundamental problems in computer vision. In this talk, I will present a unified generative statistical learning and inference framework to transfer raw images into graph representation. Based on the generic graph representation, four important issues in image and video modeling are addressed, which are photometric, geometric, dynamic and topological issues. The unified framework aims to answer questions like "What do we perceive when we look at an image or a video sequence?" It tries to identify the basic elements in images and videos, model their attributes and interactions in space and time. I will use three challenging examples to illustrate the learning and inference on these four aspects. Example 1) Textured motion modeling. Example 2) Topological changes in complex motion. Example 3) Perceptual transitions in scale-space. The integrated framework is generic. It can be widely applied to many vision applications, such as tracking, video annotation, motion segmentation, document image analysis, surveillance, and automatic cartoon animation generation.
Speaker: Yizhou Wang Title: Modeling and Inference of Complex Motion on Graphs Time: Friday(March, 04) 12:300-1:30pm Room: BH4750 Abstract As natural scenes contain a huge number of visual patterns generated by various stochastic processes, how to represent and model these diverse visual patterns, and how to learn and compute/infer those visual patterns efficiently become fundamental problems in computer vision. Motion, as an important and powerful visual cue for human perception, is one such visual pattern being studied for long, but still remains to be challenging and fundamental. In this talk, firstly, I will present a unified theory with a generative statistical learning and inference framework to model and analysis a type of challenging motion patterns called complex motion, such as falling snow, flying birds, wavy river, and dancing grass. This type of motion is characterized by the movement of a large amount of particles and wave elements, which bear rich stochastic photometric, geometric, dynamic and topologic variations. The unified framework covers all these four aspects with a generative attributed graph representation, and it attempts to answer questions like "What do we perceive when we look at a motion sequence?" This perceptual-based model is fundamental and generic. It tries to identify the basic moving elements, their attributes and interactions in space and time. It is not only good for complex motion, but also can be widely applied to many vision applications, such as tracking, video annotation, motion recognition, motion segmentation, surveillance, and automatic cartoon animation generation. The model learning and inference is achieved by employing Markov chain Monte Carlo sampling method. In the second half of the talk, I will explore the scale-space theory by treating scaling as one type of motion. The traditional scale-space theory is enriched by studying the perceptual transitions in a wide range of scales, which are explained by a set of grammar rules in the generative graph representation. The resultant multi-scale graph representation can be beneficial to many important vision applications, including image enhancement, super-resolution, feature detection, multi-resolution object recognition, and motion tracking.