We have tested our automatic modeling alogrithms on a large database of real and synthetic objects. We have also demonstrated the performance over a wide range of scales, sensors, and scene types. Finally, we have shown that automatic modeling produces accurate 3D models, limited mainly by the quality of the input data.

Here are some example automatically created models. The objects shown here were scanned with a Minolta Vivid 700 laser scanner, which also records a registered color image of each 3D view. The left-hand column for each objects shows the color image from two input views, and the right-hand column shows a rendering of the automatically constructed 3D model from approximately the same viewpoint. The VRML models are in VRML 1.0 format and can be viewed with a VRML viewer such as Cosmo player. The models have been simplified to 2,000 to 5,000 triangles to keep the file sizes small. For data sets of about 20 views, the automatic modeling process takes about 25 minutes: 5 minutes to collect data and 20 minutes to create the model.

Squirrel - 18 views - vrml model (683 KB) photograph of object automatically created model	Rabbit - 25 views - vrml model (685 KB) photograph of object automatically created model
Dwarf - 25 views - vrml model small (680 KB), large (1.7 MB) photograph of object automatically created model	My head - 34 views - vrml model (680 KB) photograph of object automatically created model

Some artifacts can be seen in the 3D models:

• A small ridge appears under the rabbit's chin. This is caused by a depth anomoly in the Vivid scanner that sometimes occurs when objects with horizontal concave ridges are scanned. Most scanners do not have this problem.
• Some of the models contain holes, which appear for several reasons. First, it is possible that part of the object was not seen from any viewpoint. This problem can usually be solved by taking additional data. Second, laser scanners have difficulty with very dark surfaces and with very shiny surfaces because the laser return is not strong enough. This can be seen in the eyes of the four figures. Finally, some laser scanners have problems near occlusion boundaries (i.e., discontinuous jumps in the range), causing the range (and the resulting 3D mesh) to be inaccurate. This problem has not been studied, and in order to eliminate this bad data, I remove the boundary vertices of each view in a pre-processing step. The input range images tend to have numerous small holes, which are enlarged by this filtering process.
• The texturing of the models is not perfect. This is due to the low-quality input images and to violations in the assumptions of the algorithm I use for creating the texture maps (the algorithm assumes Lambertian surfaces and no shadows).

Next: Large database test results