mAP (mean Average Precision) - Tech It Yourself

Hot

Thursday, 6 May 2021

mAP (mean Average Precision)

1. Introduction

mAP is a popular evaluation metric used for object detection (localisation and classification).

Object detection models such as SSD, YOLO make predictions of a bounding box and a class label.

2. True/False Positive for bounding box

For bounding box, we measure the overlap between the predicted bounding box and the ground truth bounding box IoU (intersection over union).


if the IoU value of prediction > IoU threshold, then we classify the prediction as True Positive (TF). On the other hand, if IoU value of prediction < IoU threshold, we classify it as False Positive (FP).

True or False Positive depends on 
IoU threshold. In the example above if using IoU threshold=0.2 then FP wil become TP.
In object detection:
True Positive: if True Positive for both classification and bounding box
False Positive: if otherwise

3. Calculate mAP
3.1 mAP
 
For the picture above with IoU threshold is o.5
- Average Precision (AP) is the area under the precision-recall curve.
- mAP (mean average precision) is the average of AP. 
- AP is calculated for each class and averaged to get the mAP.
The mean Average Precision or mAP score is calculated by taking the mean AP over all classes and/or overall IoU thresholds, depending on different detection challenges.

3.2 Interpolated precision


The interpolated precision, p_interp, is calculated at each recall level, r, by taking the maximum precision measured for that r.
Consider a model that predicted a dataset contains 5 apples. We collect all the predictions made for apples in all the images and rank it in descending order according to the predicted confidence level. 


The Precision at Rank 4th =  (1 + 1) / (1 + 1 + 1+ 1) = 0.5
The Recal at Rank 4th =  (1 + 1) / 5= 0.4

we smooth out the zigzag pattern:

The orange line is transformed into the green lines. We replaced the precision value for recall ȓ with the maximum precision for any recall ≥ ȓ.
We replaced all precision in [0.4:0.8] by max precision at 0.8
Pascal VOC2008 used the 11-point interpolated AP, we divide the recall value from 0 to 1.0 into 11 points — 0, 0.1, 0.2, …, 0.9 and 1.0.

AP = (1/11) * (1+1+1+1+1 +0.57+0.57+0.57+0.57+0.5+0.5)

COCO used a 101-point interpolated AP
AP75 is AP@.75 means the AP with IoU threshold=0.75
AP50 is AP@.50 means the AP with IoU threshold=5



No comments:

Post a Comment

Thường mất vài phút để quảng cáo xuất hiện trên trang nhưng thỉnh thoảng, việc này có thể mất đến 1 giờ. Hãy xem hướng dẫn triển khai mã của chúng tôi để biết thêm chi tiết. Ðã xong