mAP as a main metric for Objects Detection

Training YOLO v3 for Objects Detection with Custom Data. Build your own detector by labelling, training and testing on image, video and in real time with camera. Join here: https://www.udemy.com/course/training-yolo-v3-for-objects-detection-with-custom-data/

https://www.udemy.com/course/training-yolo-v3-for-objects-detection-with-custom-data/

• Definition
• Important terminology
• Calculating Average Precision (AP)

mAP (mean average precision) is a metric used to evaluate accuracy, in our case, for Objects Detection tasks. In general, to calculate mAP for a custom model that is trained for Objects Detection tasks, firstly, Average Precision is calculated for every class in the custom model. Then, mean of these calculated Average Precisions across all classes gives mAP. Pay attention! Some papers use Average Precision and mAP interchangeably.

Understanding the calculation process of Average Precision needs to update knowledge of definitions for used parameters.

• Threshold is used to identify whether prediction of Bounding Box (BB) can be considered as True or False. Usually threshold is set to one of the following: 50%, 75%, 95%.

• Intersection Over Union (IoU) is a measure that is used to evaluate overlap between two Bounding Boxes (BB). IoU shows how much predicted BB overlaps with so called Ground Truth BB (the one that has real object inside). Comparing IoU with threshold it is possible to define whether predicted BB is True Positive (valid in other words) or False Positive (not valid). IoU is calculated by overlapping area between predicted BB and Ground Truth BB divided by union area of two BB as shown on the Figure below.

• True Positive (TP) is a number of BB with correct predictions, IoU ≥ threshold

• False Positive (FP) is a number of BB with wrong predictions, IoU < threshold

• False Negative (FN) is a number of Ground Truth BB that are not detected

• True Negative (TN) is a number of BB that are correctly not predicted (as many as possible within an image but not overlap any Ground Truth BB); this parameter is not used for calculating metrics

• Precision represents percentage of correct positive predictions of BB (how accurate are predicted BB) and shows an ability of the trained model to detect relevant objects. Precision is calculated as following:

• Recall represents percentage of True Positive predictions of BB among all relevant Ground Truth BB and shows an ability of the trained model to detect all Ground Truth BB. Recall is calculated as following:

• Precision and Recall curve represents performance of the trained model by plotting a curve of Precisions values against Recalls values and form a kind of zig-zag graph as shown on Figure below.

In order to plot Precision and Recall curve, it is needed to collect detected BB by their confidences in descending order. Then, calculate Precision and Recall for every detected BB as it is shown in Table below. In current example, threshold is set to 50% saying that predicted BB is correct if IoU ≥ 0.5. Total number of correct predictions TP = 5 and total number of wrong predictions FP = 5.

AP is calculated by considering area under Interpolated Precision and Recall curve. Firstly, Recall values are divided into 11 points as following: [0, 0.1, 0,2 … 1] as shown on Figure below.

Then, average of maximum precision values is computed for these 11 Recall points.

From our example, AP will be calculated as following:
AP = (1/11) * (1+1+1+0.75+0.75+0.75+0.75+0.71+0.71+0.71+0.71) = 0.81