Avoid potentially spurious outlier poses during groundtruth creation

[rsoussan]

Currently our groundtruth creation pipeline maps an input bagfile (part of this process is removing low baseline images) and then registers it against an existing base surf map. The groundtruth poses are then generated by localizing the images in the input bagfile against this merged map. However, since some low baseline images are removed during map creation, not all input bagfile images exist in the resulting map. Thus, some groundtruth poses are estimates that result from matching an input bagfile image as best as possible to the merged map. This works well when the base surf map is well formed, but sometimes some "fuzzy" image locations that already existed in the base surf map can cause some spurious matches with the new bagfile. We should either add all images to the map (which isn't very feasible due to low baseline issues) or only consider the exact images from the bagfile that have been added to the groundtruth map and only generate groundtruth poses for these (perhaps with the option to interpolate for small time differences) to avoid any potential outlier groundtruth poses.

[bcoltin]

@rsoussan reword this better