PhotoModeler has been compared to other techniques of simple measurement such as using a tape measure for linear dimensions and a theodolite for 3D measurements. We have found that the accuracy of a Measurement Project is dependent on a number of factors:
• the quality of the calibration of the camera and digitizer used,
• the resolution of the camera and digitizer used,
• the geometry of the camera positions, and
• the precision with which the user marks object features as they appear in images.
For a project done with the high resolution digital camera and with reasonable user care, PhotoModeler has shown a relative accuracy in linear dimensions of around one part in five thousand (1:5,000) for man made objects (with 68% probability - the one-sigma accuracy). With higher resolution medium format metric cameras or high quality digital cameras accuracies as high as 1:30,000 have been demonstrated. Lower resolution cameras and imprecise marking can reduce the accuracy to one part in five hundred (1:500) or even lower.
Using the sub-pixel target marker greatly improves the last factor shown above (precision of marking). Given that the other factors are taken care of (good geometry, good camera calibration, etc.) one can achieve 1:30,000 or higher accuracy in a project that is all or substantially all done with sub-pixel circular target marking.
A relative accuracy of 1:5,000 means that for an object with a 10m largest dimension, PhotoModeler can produce 3D coordinates with 2mm accuracy at 68% (one standard deviation) probability.
Not all 3D coordinates that result will be of the same quality. PhotoModeler depends heavily on the user marking precise locations on the images. If a point appears in two photographs and the user guessed or approximated the location of the point in one or both of the photographs, then the resulting 3D coordinate will have low accuracy and will probably be incorrectly placed in 3D space. If there are too many points marked in this manner, other points marked precisely will be affected and the whole model will become inaccurate.