To estimate the similar places within the reference trajectory, we compute matches for every image in the trajectory to all previous images.
The match is considered valid if the cost of matching 2 images is smaller than specified non_match_cost and the images are not located nearby in the sequence. We consider the neighbourhood of 10 frames here.
Note: Other more sophisticated methods can be used for this task, like FAB-MAP2.
Non-match cost. You should roughly know the non_match_cost for a pair of images. If you know that two images should match than the non_match_cost = 1/c, where c is the cost of matching images, e.g. cosine distance between features. For cosine distance a perfect match will result in 1, so if we say that images with costs 0.7 still represent the same place, then a non_match_cost should be 1/0.7 approx. 1.4. This means that every image pair with a smaller cost will be considered as a match and all others as non-match.
The outputted file has a following format of every line:
refId1 refId2, which represents that refId1 correspondences to the same places as refId2.
If refId1 corresponds to the multiple ids, like refId3, refId4..., than there will be multiple lines representing this fact
0: refId1 refId2
1: refId1 refId3
2: refId1 refId4
...
Overestimation of similar places within reference dataset may lead to slow performance of a localizer, since a lot more place hypothesis should be checked in online phase. In general, it useful to have less false positives for this stage.
During the estimation of similar places, the binary provides the costs.txt file with all the matching cost that were computed.
We provide a script to visually inspect if the result of similar place estimation makes sense.
Use costs.txt file with simPlaces.txt file to visually inspect the matches:
python check_sim_places.py
Check the details in the script.