Rendering of a "knot-tying" data set reconstructed using a passive computer vision technique and consisting of 20k points and eight textures. The left image is rendered with traditional point splatting (one color per point), while the right one with VDTS

Abstract

We present a novel approach to render low resolution point clouds with multiple high resolution textures - the type of data typical of passive vision systems. The low precision, noisy, and sometimes incomplete nature of such data sets is not suitable for existing point-based rendering techniques that are designed to work with high precision and high density point clouds. Our new algorithm - view-dependent textured splatting (VDTS) - combines traditional splatting with a view-dependent texturing strategy to reduce rendering artifacts caused by imprecision or noise in the input data.

VDTS requires no pre-processing of input data, addresses texture aliasing, and most importantly, processes texture visibility on-the-fly. The combination of these characteristics makes VDTS well-suited for interactive rendering of dynamic scenes. Towards this end, we present a real-time view acquisition and rendering system to demonstrate the effectiveness of VDTS. In addition, we show that VDTS can produce high quality rendering when the texture images are augmented with per-pixel depth. In this scenario, VDTS is a reasonable alternative for interactive rendering of large CG models.

Keywords: Point rendering,  Picture/Image generation,  Multi-texture, Real-time

Rendering of the palm tree data set. Left: image rendered with a single color per splat; right: image rendered with VDTS

Paper

• Adobe Acrobat PDF (736 KB)

Video

• QuickTime (25.4 MB)