Calibrating cameras from internet images

Speaker

Holger Winnemöller

Date

Nov 1, 2007

Time

2:30PM to 3:30PM

Place

GRAIL (CSE 291)

Abstract

For the longest time, vector graphics representations were limited to illustrations with flat or very simple shading. Vectorized representations of photorealistic image content were largely prohibitive due to the excessive number of vector primitives required to approximate realistic image detail. Recently, several user-assisted, mesh-based methods have been introduced to address this issue.

In this work, we present an alternative vector representation and vectorization system, based on multi-scale edges, with the following advantages. Our system is fully automatic; no user-assistance is needed to guide mesh placement and sub-division. Our vector representation is intuitive; since edges occur at visible discontinuities within the image and are thus directly linked to image content, their manipulation results in predictable changes in the output image. Additionally, various edge-parameters can be controlled individually, such as location and geometry, color, and width (blur). Our vector representation is simple to manipulate; unlike mesh-based approaches, each vectorized edge is independent of any other edge. The edges, along with edge parameters, are encoded as simple splines, like those employed in most vector manipulation software. Our representation is compact; as we only store information at edges, we are not constrained by the arbitrary mesh topology of previous approaches. An integral part of our initial image edge-analysis is a computation of edge-importance. Usually, faint texture edges are less dominant than strong illumination edges. The user can choose which edges to include in the representation (or which ones to manipulate), based on their importance, thus allowing for a natural level-of-detail control and powerful compactness vs. fidelity trade-off.

In addition, our representation encodes both color and gradient information across edges. This allows us to offer various Poisson-based editing operations for vector graphics that have previously been available only in raster representations, such as texture transfer and seamless compositing. We demonstrate our work with a complete end-to-end vectorization and interactive editing solution, enabled by a real-time, GPU-based Poisson solver.

Graphics Seminar Home