Dense range data, where dense means at least one range sample per milliradian, has the potential to change our view of 3D computer graphics. Our particular area of interest has been in enhancing image-based rendering by providing color images of real scenes where each pixel has accurate range information. Thus, in addition to range image acquisition, we have also collected high-resolution color images from the same point-of-view, and built a prototype software system that builds high-resolution color photographs with depth suitable for image warping. An example of the range and color data is shown in figure 1.
The availability of range data collection has also been useful in advancing a variety of other graphics and geometry projects in our lab. For example, the collection of range data in the ``Office of the Future'' gives static information about the geometry of projection surfaces for immersive telecollaboration. As range image acquisition systems improve, dynamic range data will allow the collaborators to walk about, viewing the collaboration scene in true 3D.
A well-established research area is building geometric models from range data (point clouds and range images), and we have some new algorithms that quickly consume the range data, simplifying it to only a few percent of its original complexity.
Two additional image-based rendering applications use the dense range data. The first is our multiple-center-of-projection images where, potentially, each range sample is taken from a different, but known, location. This allows us to acquire data throughout an environment, substantially reducing occluded areas in a single data set, producing distorted images that are correctly reprojected.
The second is our work on image-based objects. Range data of an object is acquired from several locations, the images are then registered and warped to a single point-of-view. This representation has the advantage that occluded areas are drastically reduced while preserving properties of single-image warping, such as occlusion compatible ordering of rendering to preserve proper occlusion.
The final area discussed is the registration (both user-assisted and automatic) of range images taken from different locations in the same environment. One automatic technique improves upon the iterated-closest-point method  by taking into account the presence of shadows. A separate method for exploring automatic registration looks at 3D Hough transformations of separate range images.
The hardware system we have assembled is a proof-of-concept, and as such, it allows us to speculate about the future impact of high-density range images. Current hardware and algorithms in computer graphics have not considered the existence of high-density range images, and with the growing availability of similar commercial devices, new methods of handling such data need to be considered. For instance, what can be done with 100 million color samples that have positions in three dimensions? Currently, not much more than simplification, but perhaps it is time to think about new trends in rendering hardware, rendering algorithms, model-building and illumination based on high-density color and range data.
This paper describes our prototype range acquisition system, the calibration procedures, the specifics of the data collected, registration of multiple range images, the process of matching range data with color images, and the impact that it has had on our research projects. We conclude with goals for future acquisition systems, citing potential impact on graphics research.