A comparative analysis of hardware and software improvements of volume splatting E. Verg´ es , S. Grau and D. Tost May 18, 2006 Abstract This paper compares different hardware-based accelerations of the three classical splatting strategies: composite-every-sample, object-space sheet-buffer and image-space sheet-buffer. Specifically, we analyze the use of point sprites and 2D textures for the splat’s projection, the use of frame-buffer objects for the buffer composition and the use of GPGPU techniques for the transformation of voxels. We compare the efficiency of the hardware-accelerated splatting with 3D-texture mapping. In addition, we compare the rendering speed-up provided by hardware accelerations with two software-based space-leaping techniques: run-length encoding of labeled voxel models and voxel arrays. 1 INTRODUCTION The development of the last generations of graphics hardware has dramatically changed computer graphics research. Scientific visualization have gained many benefits from these advances, because they provide significant accelerations of rendering. Besides, in visualization, the ever-growing size of the datasets yields to a permanent demand of more computing power as well as hardware-based and software-based optimizations. A good example of the influence of graphics hardware technology in visualization is the current popularity of 3D-texture slicing for volume rendering. This technique consists basically of loading the entire volume in video memory and resampling it onto parallel proxy-geometry planes that are composed back-to-front to create the final image. Its speed is due to its use of the embedded trilinear interpolation hardware. Since this idea was first proposed in 1993 [CN93], the 3D-texture slicing technique has evolved and now, it incorporates volume shading [GK96], surface shading [WE98] and interactive classification [MHS99]. This evolution has been done in parallel and thanks to the graphics hardware improvements. The classical volume ray-casting can also be improved by a clever use of hard- ware capabilities. As an example, Mitsubishi’s VolumePro is a special purpose ray-casting implementation [PHK ∗ 99]. UltraVis [Kni00] is an assembler ray- casting implementation for specific types of CPUs. Finally, Kruger and Wester- 1