3.7 Process Controls

The underlying mathematics of a least-squares bundle adjustment require that a solution be found iteratively. Initial estimates of parameters are either supplied or calculated, the bundle adjustment is run to render new, refined parameter estimates, and the cycle repeats. There are two possible outcomes to an iterative process. It may converge to a solution or it may diverge and iterate endlessly. In the former case, a threshold must be defined to indicate when successful convergence has been achieved. In the latter, a control must be enforced to prevent an interminable sequence of iterations.

To define a suitable convergence criteria (which can vary from project to project) GIANT offers a convergence criteria window. Use the mouse to bring the Process Options Screen to the front of the display. The field to the lower left is named Convergence criteria, and implements this feature. Click on the window next to the label and key in an appropriate convergence criteria expressed as a percentage. The iterative cycle of the triangulation adjustment will terminate when the percentage change in the sum of squared residuals between consecutive iterations is less than the specified criteria. Note that the change can be either an increase or decrease in the sum of squared residuals. Careful thought must be given to the quality of the data when setting the convergence criteria. Image measurements from high-altitude photography measured on a low accuracy instrument, for instance, would not be able to fulfill the same convergence criteria as image measurements made with a high-precision analytical stereoviewer from low-altitude photography. The illustration below shows the convergence criteria set to 3 (%). If this field is not specified, GIANT uses a system default value of 5%.
 

 
The control against an infinite series of iterations is also found on the Process Options Screen, immediately above the Convergence criteria. The field is labeled Max number of iterations, and like the Convergence criteria has a data window that will accept the integer which represents the maximum number of iterations a solution is permitted to conduct. Guidance for setting an appropriate iteration limit must come from the size of the project and quality of the parameter estimates. Small projects may be permitted to iterate more than a large project as a general CPU-use normalizing rule of thumb. Likewise, coarse parameter estimates will likely require more iterations to refine than refined parameter estimates. It is probably judicious to be conservative in the specification of iteration limits. A triangulation can always be re-started if it appears to be convergent when the iteration limit is reached. It should be mentioned that the maximum number of iterations for a convergent solution may be reached, particularly if a Convergence criteria which is too stringent has been specified. The above screen illustrates a setting of 5 iterations maximum. If this field is not specified, GIANT uses a default value of 5 iterations.

Back to Table of Contents Back to Table of Contents

Next to Process Controls Next to "Automated Error detection facility"