Comments

Go to: [Up | Back | Table of Contents]

What Was Difficult?

Thin and small object boundaries is difficult to match correctly. More advanced matching (both similarity measure and correct-match selection) procedure is needed. Although we have developed one, it was not used, since it is roughly 100 times slower than the one we used and it was therefore not suitable for this large reconstruction.
Intensity resolution of standard TV cameras (even with AGC) is not sufficient to get high-quality images over the entire scene. We had to switch off the main light when projecting the random-texture pattern not to saturate the images and then switch it on again for taking images of natural object textures.

More precise camera calibration for a system of cameras should be developed. Calibration is the key to success in setups that use multiple physical cameras.
Proper error modeling and propagation is extremely important. If uncertainty information is available for each reconstructed point, the fusion can be done much more precisely in the sense that the density of reconstruction with the same amout of primary data can be much higher while the redundancy in the dataset can be much lower. That helps in recovering surfaces from reconstructed points.
Texture camera has to have greater resolution in intensity (at least 14 bits per channel), otherwise the images of dark regions are poor and noisy. (The illumination in an indoor scene varies by the order of several magnitudes but no gain control can be used, since the colors are to be consistent at overlaps). Ideally, points should be assigned the object albedo (and other material and surface properties) but no good method of albedo recovery is at hand for natural scenes like our office.

[Johnson & Kang 1997]: A. E. Johnson and S. B. Kang. Registration and Integration of Textured 3-D Data. Proc. Int. Conf. on Recent Advances in 3-D Digital Imaging and Modeling, Ottawa, Canada, pp. 234-241, May 1997.
[Moravec 1996]: H. P. Moravec. Robot Spatial Perception by Stereoscopic Vision and 3D Evidence Grids. Carnegie Mellon University Robotics Institute Technical Report CMU-RI-TR-96-34, September 1996.
[Weng, Cui & Ahuja 1997]: J. Weng, Y. Cui, and N. Ahuja. Transitory Image Sequences, Asymptotic Properties, and Estimation of Motion and Structure. IEEE Trans. on PAMI, 19(5):451-464, May 1997.

Last modified: Fri Nov 7 17:13:03 MET 1997