When: Wednesday, April 30, 10:00 a.m.- 11:30 a.m.

Where: 3305 Newell-Simon Hall

Xiaofeng Ren, Toyota Technological Institute at Chicago

SCS Faculty Candidate Talk

Abstract:
The grand goal of computer vision is to parse and label every perceptual structure in images. Such a complete understanding requires the use of a wide range of visual cues and the incorporation of associated processes at all levels of abstraction. I have taken an integrated approach to vision with a focus on mid-level processing, including contour/region grouping and figure/ground organization, a crucial part of visual perception that bridges together low-level signals (e.g. edges and texture) and high-level knowledge (e.g. object shape).

In this talk I will introduce a compact mid-level image representation using piecewise straight approximation of contours and the constrained Delaunay triangulation (CDT). On top of the CDT graph I will develop a unified probabilistic framework for mid-level vision, using conditional random fields (CRF) to enforce consistencies at junctions. For the first time mid-level vision is shown to be both feasible and useful, through quantitative evaluations on large human-annotated datasets. I will also demonstrate that mid-level representation and processing can apply to, and greatly facilitate, many visual tasks such as tracking objects, segmenting objects from background, and recognizing objects in both still images and videos.

<< Back