IBR-Assisted Volume Rendering

Abstract: Volume rendering at interactive frame rates remains a challenge, especially with today's increasingly large datasets. We are currently developing  a framework, using concepts from Image-Based Rendering (IBR), that decreases the required framerate for the volume renderer significantly. All the volume renderer needs to supply is a set of renderings at 'key' view points, and the IBR renderer will interpolate the intermittent frames at good accuracy. The IBR provides methods to handle both opaque and transparent datasets, and is an inexpensive process that can be run on the user's desktop machine.

Want to know more about it? Here is a paper that will be presented at the Late Breaking Hot Topics session at the 1999 Visualization conference.
 

IMAGES:

Here are intermittent IBR frames that were obtained using just one 'key frame'. Actually, a key frame is not just a single image, but consists of a set of slab images. Each of these slab images is a volume rendering of an image-aligned volume slab, enhanced with object geometry that is generated on-the-fly during the volume rendering. To demonstrate the range of views that can be obtained from one key frame, we started from the angle at which the key frame was originally obtained and then rotated out.
 

0 degs
8 degs
16 degs
24 degs
32 degs

We observe that artifacts start to appear around 16 degs. Generally, transparent datasets, such as the tomato, and dispersed datasets, such as the nerve, allow larger angular deviations than opaque, more regular objects, such as the head. Thus the IBR rendering allows the viewing angles to be varied within a cone with angle 16 degs. Hence, the user can tilt the object back and forth within a range of 32 degs. The IBR framerate was close to 30/s for all cases. In our full IBR-assisted volume rendering system, new key frames are supplied as the user rotates out of the current key frame's active range.
 

Finally, here are some quicktime clips with real-time screen captures of the IBR system at work. Again, just one key-frame was used.

UNC head

LBNL tomato dataset

Ganglion nerve cell

The underlying, on-the-fly generated geometry mesh associated with the UNC head dataset
 
 

Back to Research Homepage