While they are usually not used to evaluate object detection models, precision and recall serve as a basis to compute other metrics. A good understanding of precision and recall is, therefore, essential.
To measure precision and recall, we first need to compute the following for each image:
- The number of true positives: True positives (TP) determine how many predictions match with a ground truth box of the same class.
- The number of false positives: False positives (FP) determine how many predictions do not match with a ground truth box of the same class.
- The number of false negatives: False negatives (FN) determine how many ground truths do not have a matching prediction.
Then, precision and recall are defined as follows:
Notice that if the predictions exactly match all the ground truths, there will not be any false positives or false negatives. Therefore, precision and recall will be equal to 1, a perfect score. If a model too often predicts the presence...
