The Quantized Kd-Tree: Compression of Huge Point Sampled Models

Abstract

The ever-increasing demand for more geometric detail persists. These days, complex models can easily be obtained through the acquisition of real-world objects using 3D scanning devices. The sheer number of surface elements in these scanned datasets advocates the use of a point-based representation.
The conceptual simplicity of points scales well for large datasets and has emerged as a viable alternative over traditional primitives such as triangles and parametric surfaces. The main goal in this sketch is to compress these huge sampled models, possibly consisting of tens to hundreds of millions of points.
Even though using points eases storage requirements (e.g. by not storing connectivity information), compression is likely to be required for real-world applications. A scene of 80M points for instance, already requires 1GB of memory just for storing 3D positions at 32 bit precision, which does not even include normals and colors. Compression enables us to render such and even larger objects without requiring expensive disk access.