Term |
Description |
Example |
Precision |
Precision is a measure of usefulness as it measure the rate of true positives (objects detected are actually vessels) |
If one out of every 4 vessels detected is actually a ship, our precision is 25% |
Recall |
Recall is a measure of completeness as it measures the rate of false positives (objects that are missed) |
If one out of ever four actual vessels are detected, our recall is 25% |
Average Precision (AP) |
Average Precision (AP) is the standard object detection metric used by the computer vision community. The metric incorporates precision, recall and how confident the model is in each detection. Average Precision is calculated independently for each class |
|
mAP (Mean Average Precision) |
Mean Average (mAP) can be used in multi-class problems and is calculated by taking the mean AP over all classes. |
|
F1 Score |
F1 is a balance between precision and recall |
|
IoU (Intersection over Union)
|
This is a measurement of accuracy. How well the actual ship object fit inside the bounding box. This is computed by taking a ratio of overlap between the two bounding boxes. |