This is a list of tips for getting the best results out of 3dpalign.
User Interface
-
Mouse button use is in general designed around the following rules: The left mouse button only manipulates your view of the world (e.g. rotating a view of a range image, or moving an icon) but does not make permanent changes to the information you are inputting. The center mouse button actually does work: selecting images, identifying features, and so forth. The right mouse button is used only for context-dependent menus. These general guidelines should help you remember how to operate the user interface.
-
Resist the urge to close the aligned view windows (pairwise, full global and fast global) using the window manager's widgets (such as the close box in the upper left on Motif systems). Doing so will cause the program to abort. Instead, close these windows using the commands in the "Window" menu.
Feature identification
-
You are only required to "connect" (identify corresponding features between) enough pairs of range images such that every range image is somehow connected to the whole. However, for best results, it's good to identify corresponding features in ALL overlapping pairs of range images, even when it seems redundant. 3dpalign makes use of all the extra information and can generate a better fit.
-
The minimum number of features necessary to "connect" two range images is three. However, you will find that you will get better results with more features. It is recommended that you use at least five. Furthermore, you will get better results if these features are spread out fairly evenly over the entire region of overlap.
-
The "strength" of the connection between two range images is determined by the number of feature correspondences connecting them. Therefore, if it is important that two particular range images match very well in the final output, you should identify a large number of corresponding features between those two range images.
Full global alignment
-
"Full global alignment" (or refined alignment) is not guaranteed to improve the alignment. In some cases, repeated application of full global alignment may actually worsen the subjective appearance of the aligned range images. Therefore, you should normally work with your feature correspondences until the "fast global alignment" looks as good as possible FIRST; then attempt to use full alignment later only if it appears to be improving the appearance of your aligned range images.
-
"Full global alignment" (or refined alignment) requires that your calibration data (the various coefficients and transforms in your *.params files) is reasonably accurate, because calibration information is used to optimize the process of attempting to automatically find feature correspondences. If your calibration data is inaccurate, then full global alignment may not work very well; or, you may even get error messages stating that it is unable to work at all.
-
"Full global alignment" (or refined alignment) works best if your data is very clean. If you have noisy data, or your data contains noticeable distortions, you will typically find that you will get better results using "fast global alignment."
Daniel Azuma
(dazuma@kagi.com)
Last updated 28 August 1997