Fragment-History Volumes

TL;DR

Fragment-History Volumes (FHV) are a novel GPU-based technique that records all rasterized fragments to enable view-independent rendering, fast resampling, and potential global illumination evaluation in real-time graphics.

Contribution

The paper introduces Fragment-History Volumes, a new method for storing rasterization data that supports view-independent rendering and advanced lighting effects on current hardware.

Findings

FHV enables fast, view-independent image resampling.

FHV supports complex lighting and material evaluations.

FHV can be stored and processed efficiently on GPU.

Abstract

Hardware-based triangle rasterization is still the prevalent method for generating images at real-time interactive frame rates. With the availability of a programmable graphics pipeline a large variety of techniques are supported for evaluating lighting and material properties of fragments. However, these techniques are usually restricted to evaluating local lighting and material effects. In addition, view-point changes require the complete processing of scene data to generate appropriate images. Reusing already rendered data in the frame buffer for a given view point by warping for a new viewpoint increases navigation fidelity at the expense of introducing artifacts for fragments previously hidden from the viewer. We present fragment-history volumes (FHV), a rendering technique based on a sparse, discretized representation of a 3d scene that emerges from recording all fragments that pass the rasterization stage in the graphics pipeline. These fragments are stored into per-pixel or per-octant lists for further processing; essentially creating an A-buffer. FHVs using per-octant fragment lists are view independent and allow fast resampling for image generation as well as for using more sophisticated approaches to evaluate material and lighting properties, eventually enabling global-illumination evaluation in the standard graphics pipeline available on current hardware. We show how FHVs are stored on the GPU in several ways, how they are created, and how they can be used for image generation at high rates. We discuss results for different usage scenarios, variations of the technique, and some limitations.