Refining 3D Reconstructions: A Theoretical and Experimental Study of the Effect of Cross-Correlations

Dealing with Noise in Multiframe Structure from Motion

“…These results were previously reported in [52], except for the results of the Blind Averaging algorithm. They are included here for completeness.…”

“…Of the four experiments, the first reported here is identical to one reported in [52]; however, we have added the Blind Averaging results in this paper. The image sequence used for the second experiment was also used for experiments in [52].…”

“…The image sequence used for the second experiment was also used for experiments in [52]. However, as described below, certain aspects of the experiment on this sequence in the current paper are new.…”

“…Subsequent to our first description and experimental testing of this algorithm [34,49,50,52], several algorithms along similar lines have been reported [5,39].…”

“…For instance, if the displacement of the camera between two images is identified incorrectly, then the estimated positions of all points (based on these images) will be displaced away from their actual ones in approximately the same direction; their position errors will be correlated. Conversely, the correlations in the structure errors are the record of the motion error [49,34,52]. In order to include the potentially large effects of motion error and to compensate for them, a record of these correlations must be maintained in the current error estimate.…”