| Research Abstract |
We have been developing methods to automatically generate virtual-reality models through observation of real objects. This research spans both geometric and photometric aspects of modeling.
As for geometric aspect, we have developed a method to obtain shapes of transparent objects such as glasses or jewelry using polarization. Usual shape-determination methods, such as laser range finders or binocular stereo, are only applicable to opaque surfaces ; few methods can be applied to transparent surfaces. Transparent surfaces have highlight and the degree of polarization in highlight depend on surface orientations. The degree of polarization is a double-value function in visible light ?one polarization value corresponds to two surface orientations -- while it is a single-value function in infrared light. On the other hand, accuracy of measurement is much higher in visible light than in infrared light. Thus, by using both infrared and visible light, we have developed an algorithm to determine surface orientations of transparent surfaces. We have conducted experiments to evaluate the accuracy of the method and demonstrated its effectiveness.
As for photometric modeling, we have developed an eigen-texture method. This method stores various views of an object *er a geometric model. These views are compressed with respect to the coordinate defined over the object surface. Then, from these compressed views, the method can re-generate arbitrary views. Since views are compressed on the object Coordinates, each part of view are highly correlated, and, thus, the method can achieve very high compression ratio. We have implemented this method, and conducted various experiment to evaluate the performance of this method.
We have published these results on international journals and presented those at international conferences.
|
