The performance of the place recognition algorithm is typically evaluated according to precision, recall metrics, and precision-recall curve. The matches consistent with the ground truth are re- garded as true positives (TP), the matches inconsistent with the ground truth are false positives (FP), and the matches erroneously discarded by the place recognition algorithm are regarded as false negative matches (FN). Precision is the proportion of matches recognized by the place recognition algorithm as true positives matches, and recall is the proportion of true positives matches to the total number of actual matches in the ground truth, that is,
Precision = T P /T P + F P ,
Recall = T P/ T P + F N .
In addition, for place recognition, the maximum recall at 100% precision denoted as R P 100 is also an important performance in- dicator; R P 100 is widely used to evaluate place recognition algo- rithms and is also used in our subsequent experiments. That is because a single false positive may cause irremediable failures for many robotic applications. For example, in the context of loop clo- sure detection, false positives are graver than false negatives [25] , although this criterion may cause the algorithms to reject several exact matches in the ground truth. Thus, to determine the robust- ness of place recognition, the recall rate at the precision of 100% is a good indicator.
via: Sequence searching with CNN features for robust and fast visual place recognition